Improving Identity-Robustness for Face Models
Extracted source text · arXiv 2304.03838
\documentclass[10pt,twocolumn,letterpaper]{article}
\usepackage{iccv}
\usepackage{times}
\usepackage{epsfig}
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{algorithm, algorithmic}
\usepackage{booktabs}
\usepackage{placeins}
\usepackage{multirow}
\def\x{{\bf{x}}}
\def\y{{\bf{y}}}
\def\z{{\bf{z}}}
\def\w{{\bf{w}}}
\def\p{{\bf{p}}}
\def\R{{\mathbb{R}}}
\def\D{{\mathcal{D}}}
\def\B{{\mathcal{B}}}
\def\N{{\mathcal{N}}}
\def\I{{\mathcal{I}}}
\def\U{{\mathcal{U}}}
% Include other packages here, before hyperref.
% If you comment hyperref and then uncomment it, you should delete
% egpaper.aux before re-running latex. (Or just hit 'q' on the first latex
% run, let it finish, and you should be clear).
\usepackage[pagebackref=true,breaklinks=true,letterpaper=true,colorlinks,bookmarks=false]{hyperref}
%\cvprfinalcopy % *** Uncomment this line for the final submission
%\iccvfinalcopy
% \def\iccvPaperID{7562} % *** Enter the CVPR Paper ID here
\iccvfinalcopy
\def\httilde{\mbox{\tt\raisebox{-.5ex}{\symbol{126}}}}
% Pages are numbered in submission mode, and unnumbered in camera-ready
%\ifcvprfinal\pagestyle{empty}\fi
\begin{document}
%%%%%%%%% TITLE
%\title{Identity Proxy Neighbourhood for Face Models}
%\title{Talent fairness enhancement by proxy structure exploration}
%\title{Robustness across identitues using Distributionally Balanced Sample Weighting}
\title{Improving Identity-Robustness for Face Models}
% \title{Training Identity-robust Face Models using Distributional Class Balancing}
\author{Qi Qi\thanks{Most of the work was done while the author was an intern at Netflix. The first version of the paper was released at April 7th, 2023.}\\
The University of Iowa\\
Iowa City, IA\\
{\tt\small qi-qi@uiowa.edu}
% For a paper whose authors are all at the same institution,
% omit the following lines up until the closing ``}''.
% Additional authors and addresses can be added with ``\and'',
% just like the second author.
% To save space, use either the email address or home page, not both
\and
Shervin Ardeshir\\
Netflix\\
Los Gatos, CA\\
{\tt\small shervina@netflix.com}
}
\maketitle
%\thispagestyle{empty}
%%%%%%%%% ABSTRACT
\begin{abstract}
Despite the success of deep-learning models in many tasks, there have been concerns about such models learning shortcuts, and their lack of robustness to irrelevant confounders. When it comes to models directly trained on human faces, a sensitive confounder is that of human identities. Many face-related tasks should ideally be identity-independent, and perform uniformly across different individuals (i.e. be fair). One way to measure and enforce such robustness and performance uniformity is through enforcing it during training, assuming identity-related information is available at scale.
However, due to privacy concerns and also the cost of collecting such information, this is often not the case, and most face datasets simply contain input images and their corresponding task-related labels. Thus, improving identity-related robustness without the need for such annotations is of great importance. Here, we explore using off-the-shelf face-recognition embedding vectors, as proxies for identities, to enforce such robustness. We propose to use the structure in the face-recognition embedding space, to implicitly emphasize rare samples within each class. We do so by weighting samples according to their conditional inverse density (CID) in the proxy embedding space. Our experiments suggest that such a simple sample weighting scheme, not only improves the training robustness, it often improves the overall performance as a result of such robustness. We also show that employing such constraints during training results in models that are significantly less sensitive to different levels of bias in the dataset.
\end{abstract}
%%%%%%%%% BODY TEXT
\vspace{-0.05in}
\section{Introduction}
\vspace{-0.05in}
%% models are often not robust/fair
%% why does bias occurs?
%% Why robustness for face models is specifically important?
%% How to mitigate bias
%% with groups
%% without groups
%% how do we do it
%% Why robustness for face models is important
Given the success of machine learning models, and their deployment at scale, having a more extensive evaluation of the robustness of such models is of utmost importance. Given the nature of training such models, there is always the potential for these models to rely on irrelevant and spurious shortcuts. Relying on such shortcuts could have immense negative consequences when the dataset and tasks are defined around humans. A prevalent type of such datasets and tasks are those defined on human faces, ranging from regression tasks such as estimating pose\cite{albiero2021img2pose}, facial-landmarks\cite{wu2019facial}, etc, to classification tasks such as facial-expressions classification\cite{huang2019facial}, and generative tasks such as avatar creation\cite{alldieck2018detailed}, etc. A common attribute of many of such face-centric tasks is the fact that the model performance, should be identity independent by definition, yet this aspect of a model is often not taken into account during training and evaluation. Two models trained on a face-related task can have similar overall performance, but very different levels of robustness across different individuals. The toy example in Figure~\ref{fig:face_model_illstration} illustrates this concept. This disparity in performance often gets baked into the model due to bias in the training data, as data points belonging to different subpopulations may have a different level of class imbalance. More specifically, if person 1 smiles 90\% of the time, and person 2 smiles 10\% of the time, a smile classifier can easily latch on to the facial features of person 1 as a shortcut to reduce training loss significantly. Thus, it could always label images of person-1 as smiling, because of the person's identity, and not the facial expression.
Awareness of identity/group labels would allow for mitigation approaches to prevent such bias, such as recent efforts in adversarial training~\cite{zhang2018mitigating, elazar2018adversarial}, model interpretation method \cite{rieger2020interpretations} and objective regularization~\cite{bechavod2017penalizing}, which aim to reduce the disparity between different groups using the ground-truth group labels $g\in G$. In many practical scenarios, however, such information is not available at scale during training and evaluation. Also, collecting such detailed annotation could be costly and undesirable due to three main reasons: First, annotating every sample data point with all their potential types of group-membership information could be extremely costly. Second, collecting and maintaining such detailed categorical labels on human faces raise data-privacy concerns. And third, the nature of many types of such group memberships may be extremely subjective. In addition to the previous hurdles in obtaining such data, most current large-scale datasets, lack such annotations at scale, which is another testament to the need for approaches that are not reliant on the availability of such additional information. As a result, improving fairness when the ground-truth group labels $g\in G$ are unknown is of utmost importance, and has given rise to an area of research often referred to as "fairness under unawareness". When it comes to "fairness under unawareness" for face models, the only earlier work is ~\cite{ardeshir2022estimating} which aims to measure the performance disparity of a model in the absence of group information. A disparity method (Disparity across Embedding Neighborhoods) is proposed, which approximates Rawlsian Max-Min (RMM) across groups $g\in G$, solely based on face-recognition embedding vectors. The neighbors of a sample are defined as the samples whose euclidean distance in the face-recognition embedding space is less than a predefined threshold. The aforementioned work solely focused on approximating disparity for a given model. In this work, however, we focus on using such intuition to reduce such disparity during training and directly optimize for such an objective. In other words, given a face dataset and solely its task labels, and without any group information, we explore if we can use embeddings from an off-the-shelf face recognition model to reduce the performance disparity of such a model across different individuals.
\begin{figure*}[t]
\centering
\includegraphics[width = \linewidth]{./figures/CID_toy_example.png}
\caption{Toy example visualizing our proposed approach. The task is predicting if a face image is smiling (green) or not (red). The Biased Classifier shows how a biased dataset could lead to a model latching on to spurious features (identity) for an identity-independent task (smiling). We propose extracting face-recognition embeddings and using the structure in that space to weight rare samples within each class. More specifically, for each class (green or red), each sample is weighted based on its class-conditioned inverse density in the proxy (face recognition) embedding space. As a result, in each class, the rare samples are emphasized in the Robust Classifier.}
\label{fig:face_model_illstration}
\end{figure*}
% \begin{figure}[t]
% \centering
% \includegraphics[width = 0.7\linewidth]{./figures/models_for_id.png}
% \caption{The toy example of predicting the color for each sample. Hence the task label $Y\in \text{\{Red, White\}}$. And the samples include three groups $G\in\{\text{Triangle, Circle, Rectangle}\}$. The two models have the same overall accuracy of 0.8, however, the accuracy for different groups varies. The group accuracy of $\{g_1, g_2, g_3\}$ are $\{1, 0.4, 1\}$ in the first model and $\{0.8, 0.8, 0.8\}$ in the second model. When it comes to the face models, the color maps whether a person is smiling or not, and each shape represents an identity.}
% \label{fig:face_model_illstration}
% \end{figure}
% %% How do we address it
% In this paper, we focus on models that are trained on human faces, which have wide applications across a variety of fields, including entertainment~\cite{herruzo2019see}, surveillance~\cite{baharani2019real}, and medicine~\cite{alissa2021parkinson}. Among them, face recognition models~\cite{prasad2020deep} are trained by encouraging image embeddings from the same person to be close to each other, while embeddings from different individuals are further away. Hence, the embedding space of an off-the-shelf face-recognition embedding would contain facial features resulting in recognizing an individual. Assuming its generalization, a neighborhood in such an embedding space could serve as a lookalike model, as faces with similar facial features should have closer embeddings. Therefore, many sensitive attributes, such as demographic information are implicitly contained in such embeddings. This concept was quantified in \cite{ardeshir2022estimating}, where a disparity method (Disparity across Embedding Neighborhoods) was proposed, which approximates Rawlsian Max-Min (RMM) across groups $g\in G$. The neighbors of a sample are defined as the samples whose euclidean distance in the face-recognition embedding space is less than a predefined threshold. The aforementioned work solely focused on approximating disparity for a given model. In this work, however, we focus on using such intuition to reduce such disparity during training and directly optimize for such an objective. In other words, given a face dataset and solely its task labels, and without any group information, we explore if we can use embeddings from an off-the-shelf face recognition model to reduce the performance disparity of such a model across different individuals.
As identity bias is often induced due to class imbalance within a subpopulation of a dataset, we propose to exploit the structure of the training data points in the face-recognition embedding space and enforce class balance for any subpopulation that shares similar facial features. This is achieved by weighting the samples according to their conditional inverse density (CID) in the proxy embedding space, resulting in equalizing the effect of task-positive and task-negative labels for each local neighborhood in the proxy embedding space. In other words, we aim to equalize the total weight for positive and negative samples for people whose facial features look like any arbitrary embedding $z$. Our experiments show that such a simple sample weighting scheme not only reduces the performance disparity of the trained model across different individuals and groups but also makes the training more robust to distribution shifts between train and test, and different levels of dataset bias. We evaluate such robustness by designing a stress test where we artificially manipulate the bias in the dataset, and control its level of bias.
% We are focusing on the face models and face images. Hence, demographic information, such as gender, ethnicity, and identity, are natural sensitive groups that worth attentions.
% artwork generation in streaming industry[], person reidentification in video surveillance[].
\vspace{-0.1in}
\section{Related Work}
\noindent
The bias mitigation methods can be broadly categorized into two groups based on the availability of group information, denoted by $G$. When $G$ is available, the proposed methods usually explicitly incorporate such group information into the training process to reduce bias, such as penalizing the group difference as a regularizer~\cite{bechavod2017penalizing,beutel2019fairness}, enhancing fair representations via contrastive learning only using task-relevant features by constructing hard negative pairs from different groups~\cite{zhangfairness, park2022fair, du2021fairness}, learning an adversary head to reduce models' ability to distinguish group-relevant features that amplify biases~\cite{zhang2018mitigating, wadsworth2018achieving}, and so on.
Our work can be categorized under the "fairness under unawareness" umbrella, where group information $G$ is unavailable during training. Due to the lack of group information, improving model robustness to minimize spurious correlations has become one of the mainstream methods to enhance model fairness. This can be achieved through various methods such as invariant risk minimization~\cite{arjovsky2019invariant,adragna2020fairness}, distributionally robust optimization~\cite{ben2013robust, caton2020fairness, sagawa2019distributionally,hashimoto2018fairness,li2021tilted, qi2020attentional, lahoti2020fairness}, and class balancing methods~\cite{yan2020fair,cui2019class,huang2016learning, wang2017learning}.
\noindent
\textbf{Invariant Risk Minimization (IRM)}
IRM~\cite{arjovsky2019invariant} was proposed to address the domain-shifting problems by learning invariant feature representations that have good generalization ability between training and testing. To achieve this,
IRM optimizes the losses across different data distributions to improve model robustness.
~\cite{adragna2020fairness} shows the effectiveness of IRM in reducing the disparity between different protected groups in the toxicity classification task on Civil Comments natural language process dataset when $G$ is unavailable.
%on the Civil Comments natural language process dataset.
\noindent
\textbf{Distributionally Robust Optimization (DRO)} Compared with ERM which minimizes the average sample losses, DRO aims to focus on the largest errors~\cite{ben2013robust, caton2020fairness} by assigning robust weights $p_i$ to samples proportional to the loss scales.
%could be redundant, no need for so many formulas in related work.
%$\min_{\w}\max_{\p\in \U}\sum_i p_i\ell_i$, where $p_i$, proportional to losses scale $\ell_i$ denotes the robust weight of a sample, and $\U$ denotes the distribution constraint of $\p$ to obtain a valid robust weights such that $0< p_i <1,\sum_ip_i = 1$.
To verify the validity of DRO in improving fairness under unawareness,
%When it comes to fairness unawareness,
~\cite{liu2021just} empirically shows that most of the samples with the largest errors belong to the worst group, ~\cite{hashimoto2018fairness} theoretically proves that DRO can control group disparity amplification in every iteration, and~\cite{li2021tilted, qi2020attentional} shows the effectiveness of DRO in different fair applications.
Recently, ~\cite{lahoti2020fairness} proposed a generalized DRO method, namely Adversarially Reweighted Learning (ARL), by parameterizing the robust weight $p_i$ with an adversarial network $\phi$ and leveraging the concept of a computationally-identifiable subgroup of largest errors~\cite{hebert2018multicalibration} to improve model fairness.
% In the experiments, we implement the DRO method proposed in \cite{li2021tilted, qi2020attentional}, and ARL~\cite{lahoti2020fairness} for a fair comparison.
\noindent
\textbf{Class Balancing}~\cite{yan2020fair} proposed a cluster-based balancing method by generating the minority samples for each cluster using the upsampling K-Means SMOTE~\cite{last1711oversampling}, however, this upsampling clustering-based method only applicable to small tabular data and lead to excessive training time for larger datasets. Alternatively,
class balancing reweighting methods~\cite{cui2019class,huang2016learning, wang2017learning} are widely used to improve model robustness in large datasets by assigning weights to balance the contribution of different classes. These weights of each sample are typically inversely proportional to the number of class samples, which helps to improve the performance of minority classes and alleviates the spurious correlations between the sensitive groups and classes incurred due to the lack of samples in minority class.
To summarize, the IRM, DRO, and class-balancing methods rely on the implicit preservation of sensitive group (categorical) information in model predictions, either through losses~\cite{adragna2020fairness, ben2013robust, caton2020fairness, sagawa2019distributionally,hashimoto2018fairness,li2021tilted, qi2020attentional} or feature representation embeddings~\cite{lahoti2020fairness,yan2020fair}. However, in our case, this assumption does not hold as we are focusing on face-centric tasks that are defined identity independent. Therefore, our CID method resorts to face recognition embeddings to have better group proxies. In the supplementary, we provide comparisons and draw parallels between our approach and the aforementioned methods under certain assumptions.
% To summarize, the key assumption for the aforementioned IRM, DRO and class-balancing methods is that the sensitive group (categorical) information is implicitly preserved to some extent in task-related model predictions, either through losses~\cite{adragna2020fairness, ben2013robust, caton2020fairness, sagawa2019distributionally,hashimoto2018fairness,li2021tilted, qi2020attentional} or feature representation embeddings~\cite{lahoti2020fairness,yan2020fair}, which is not held in our case as the task by definition unrelated to identity. Hence, in our CID method, we resort to face recognition embeddings to have better group proxies. We draw parallels and comparisons between our approach and the aforementioned under certain assumptions in the supplementary.
% by assigning weights that balance the contribution of different classes.
%Before we ending
%Class reweighting methods that assign class-balancing robust weights that are inversely proportional to the number of class samples to balance the contributions of different classes and improve the performance of the minority classes are widely used to improve model robustness in large datasets~\cite{cui2019class,huang2016learning, wang2017learning}.
%Hence, the spurious correlations between the sensitive groups and classes incurred by the lack of samples from the minority classes will be alleviated.
% in each cluster feature representations.~\cite{yan2020fair} is the most closet work with us that shares a similar idea that enhancing class balancing in the local area (cluster), however
% Instead, we are focused on face images and make use of off-the-shelf face recognition model representations to improve model fairness by assigning robust weights to samples.
% It shares a similar idea with our work, however, the key difference lies in it is limited to the smaller dataset due to the overwhelming training time.
% To improve model robustness, class-level reweighting methods balance the contributions of different classes by assigning class-level weights to samples. Typically, the class-level weights are designed to be inversely proportional to the number of class samples~\cite{cui2019class,huang2016learning, wang2017learning}. Hence, the samples in the minority classes will be up-weighted to improve the worst-case performance, which alleviates the spurious correlations between the sensitive groups and classes due to the lack of samples from the minority classes.
% \noindent
% \textbf{Robust Loss}
% The robust loss is another research direction for handling data imbalance problems. For example, focal loss~\cite{lin2017focal}, LDAM~\cite{cao2019learning}.
% Recently,~\cite{qi2022fairness} proposed an adversarial attribute robust loss to reduce the disparity across different sensitives when the group information is available. In this paper, we minimize a robust loss derived from the conditional inverse density min-max objective defined on the proxy embedding vectors when $G$ is not known.
%\vspace{-0.1in}
\section{Approach}
\label{sec:approach}
%\vspace{-0.1in}
Given a dataset of images of faces, and an identity-independent face-related task such as predicting a facial expression (e.g. smiling), we aim to train a classifier that performs robustly across face images of different people. We refer to training labels related to the task of interest (smiling), as \textit{task labels}. We assume that such labeling (whether a face is smiling or not) is given to us for training and test set. On the contrary, we assume that no \textit{identity} label is given to us during training. Identity labels specify which images belong to which person (person-1, person-2, ...), across which we would like to enforce fairness/robustness.
We also assume that we have access to an off-the-shelf face-recognition model, using which we can extract an embedding for each face image. Our goal is to train a model for that task, that performs robustly (fairly) across different individuals on the test set. Please note that in our experiments, we solely use the identity-labeled test sets to validate the robustness of our approach, and we do not use such labels during training. Formally, given a dataset $\D = \{X\times Y\} = \{(\x_i,y_i )\}_{i=1}^{|\D|}$ with size $n=|D|$, the total number of classes $C$, i.e, $|Y| = C$. $\D_{y} = \{(\x_i,y_i)|y_i = y, i\in [1,\cdots,|\D|]\}$ represents the samples whose task label is $y\in Y$. $g_i\in G$ denotes the identity/group that sample $i$ belongs to, across which performance disparity should be mitigated. Under our setup, group/identity labels $G$ are unavailable during training. Instead, the embedding vectors $\{\z_i\}_{i=1}^{|\D|}$ are extracted from a face recognition model and are provided as proxies for the group/identity membership.
% \textcolor{red}{Compared with existing bias mitigation settings when the group labels are unavailable, we claim that our settings are more changeling due to the fined-grained groups, i.e, thousands of identities. In addition, how to evaluate the improved fairness across different identities also matters. In the following, we proposed an effective fine-grained sample weight method, and introduce the AUMM curves that extended from the Rawlsian Min-Max metric as a robust evaluation metric. }
\vspace{-0.05in}
\section{Training}
\vspace{-0.05in}
\label{sec:training}
Inspired by recent efforts in ~\cite{diana2021minimax,hashimoto2018fairness,lahoti2020fairness}, we define our objective as a min-max form, which encourages emphasis the performance of the model on the least accurate areas of the embedding space:
\vspace{-0.05in}
\begin{align}
\label{eqn:obj}
\min\limits_{\w} \sum\limits_{i=1}^n& \frac{p_i^\tau}{Z_{y_i}}\ell(\w;\x_i,y_i) \\
\label{eqn:robust_fairness_constraint}
\text{s.t} \arg\max\limits_{\p_i\in\Delta_{\D_{y_i}}} \sum\limits_{j\in \D_{y_i}} &p_{ij}\z_i^\top\z_j - \tau \text{KL}( \p_i, \frac{\textbf{1}}{|\D_{y_i}|})
%p_i^\tau \in \arg\max\limits_{\p_i\in\Delta_{\D_{y_i}}} \sum\limits_{j\in \D_{y_i}} &p_{ij}\z_i^\top\z_j - \tau \text{KL}( \p_i, \frac{\textbf{1}}{|\D_{y_i}|})
\vspace{-0.2in}
\end{align}
% \textcolor{red}{$\frac{1}{|\D_{y_i}|}\sum\limits_{j\in\D_{y_i}}\z_i^\top\z_j$
% measures the average pairwise similarity between the sample $i$ and $\D_{y_i}$. The larger the value, the denser the class distribution, and vice visa. To take the local neighborhood structure into consideration, we emphasize more on the large similarities by penalizing the uniformity pairwise similarity value. The larger the matter, the higher density in its neighborhood determined by $\tau$}.
where $p^\tau_i:= p_{ii}$ denotes the sample weight, $\ell(\w;\x_i,y_i)$ denotes the prediction loss, and $Z_{y_i} = \sum_{i\in \D_{y_i}}p^{\tau}_{i}$ is the class-level normalization parameter to guarantee each class contributes equally. To obtain $p_{i}^\tau$, the maximum constraint in (\ref{eqn:robust_fairness_constraint}) is imposed on the pairwise similarity of proxy embedding vectors, leveraging the proxy neighborhood structure associated with each sample. To be more specific, for $\forall (\x_i,y_i)\sim \D$, $\p_i= (p_{i1}, \cdots,p_{ii},\cdots, p_{i|\D_{y_i}|})$ refers to the weight assigned to each sample based on $\{\z_i^\top\z_j \}_{j\in\D_{y_i}}$ and satisfies $\Delta_{\D_{y_i}}:=\{\sum_{j} p_{ij} = 1, p_{ij}\geq 0\}$. The KL divergence regularizer $\sum_jp_{ij}\log(|\D_{y_i}|p_{ij})$ between the uniform distribution $1/|\D_{y_i}|$ and the pairwise weights $\p_i$
encourages the model to focus on the local neighborhood. The regularizer hyperparameter $\tau$ measures the proximity and magnitude of the neighborhood, which will be explained in next section.
% Maximize $\p_i$ for each sample encourages the model focusing on the hard samples, which will explain in the following sections.
% The robust weight $p_{ii}$ is able to improve the model robustness during the training and reduce the ground truth disparity by exploring the neighbourhood structure of every sample.
% are designed for each sample to approximate the ground truth group information in the embedding space,
% i.e, the pairwise robust weights $p_{ij}$ is defined on top of the
% ie robust soft assignment neighbourhood method that considers both neighbourhood similarity and point densities where $p_{ij}$ denotes the importance of a sample $j$ in the neighbourhood $N_i$. Then we keep the $p_{ii}$ for the sample $i$.
\subsection{Batch-wise Implementation using Conditional Inverse Density (CID)}
\label{sec:rpse}
Here we explain how we practically optimize the objective mentioned above using a sample-weighting scheme based on the conditional inverse density (CID for short) of each datapoint in the proxy embedding space.
To expand, we consider the practical batch-wise training scheme such that the constraint set $\D_{y_i}$ in Eqn~(\ref{eqn:robust_fairness_constraint}) is defined as the samples having the same task labels in the current batch $\B$, i.e, $\B_{y_i}$ (thus, conditioned on task label).
Thanks to the strong concavity of $\p_i$ in~(\ref{eqn:robust_fairness_constraint}) and the specific structure of KL divergence, the close form solution of $p_i^{\tau}:= p_{ii}$ is obtained by taking the first derivative of $\p$ in (\ref{eqn:robust_fairness_constraint}) equals to 0, i.e,
\vspace{-0.1in}
\begin{align}
\label{eqn:p_ii}
p_{i}^\tau = \frac{\exp(\frac{\z_i^\top\z_i}{\tau})}{ \sum\limits_{k=1}^{|\B_{y_i}|}\exp(\frac{\z_i^\top\z_k}{\tau})}
\end{align}
\vspace{-0.1in}
where the numerator is the exponential of the inner product of the proxy embedding vector $\z_i$ of sample $(\x_i, y_i)$. The denominator explores the neighborhood
proxy structure by aggregating of the exponential pairwise similarities of proxy vectors between sample $(\x_i, y_i)$ and $\B_{y_i}$.
Even though the constraint set is defined in $\B_{y_i}$, the skewness property of exponential function $\exp(\cdot/\tau)$ for large similarities pairs encourages the denominator to focus on the local neighbors of $(\x_i, y_i)$ that share the same facial features.
$p_i^{\tau}\in (0, 1]$ represents the importance of the sample $(\x_i, y_i)$ in the local neighborhood. The fewer the samples in the local neighborhood, the higher the $p_i^\tau$. Hence, $p_i^\tau$ is inverse proportional to the class-conditional sample density in the local neighborhood and emphasizes on the rare samples within each class.
In~\cite{ardeshir2022estimating}, the performance of a model across different local neighborhoods in the proxy embedding space is used to estimate disparity across identities/groups. Hence a local neighborhood could be seen as an approximation for a subpopulation/group/identity $g_i$. ~\cite{ardeshir2022estimating} also illustrates that there are different neighborhood sizes that better approximate different group memberships. To capture the same concept, in our formulation
$\tau$ controls the skewness of the exponential function, which influences the size of the local neighborhood. Thus, we fine-tune the hyper-parameter $\tau$ to allow for exploring different neighborhood sizes and therefore different density estimations. Figure~\ref{fig:soft_assignment_weights} shows the impact of $\tau$ on the weight of three different samples. As it can be seen, as $\tau \rightarrow \infty$, the weights converge to $p^\tau_{i} \rightarrow \frac{1}{|\B_{y_i}|}$ which is simply the inverse of per-class frequency.
% %%%======
% To express the influence of $\tau$, we consider a normalized proxy vector batch $\B$,
% i.e, $0\leq \| \z_i\|\leq 1, \forall i\in\B, \text{and}\ \z_i^\top\z_i = 1$, and denote $1/p_i^{\tau}$ as the size of the soft local neighborhood. Hence, $\tau$ can be viewed as a radius parameter that maps to the size of the neighborhood. $\forall (\x_i,y_i)\in \B$, according to~(\ref{eqn:robust_fairness_constraint}), $p^\tau_{i} = 1$ when $\tau=0$. Then the neighborhood of $\z_i$ only includes itself and $Z_{y_i} = B_{y_i}$. When $0<\tau <\infty$, the larger the $\tau$, the larger the neighborhood radius, the smaller the $p_i^\tau$.
% When $\tau \rightarrow \infty$, according to Eqn~(\ref{eqn:p_ii}), $p^\tau_{i} \rightarrow \frac{1}{B_{y_i}}$ and $Z_{y_i} = 1$.
% When $\tau = 0$ and $\tau \rightarrow \infty$, normalized by $Z_{y_i}$, the weight $p_i^\tau/Z_{y_i}$ is proportion to $1/B_{y_i}$ in (\ref{eqn:obj}). Hence, the proposed CID method boils down to the batch wise inverse frequency class reweighting by just making use of the class information. When $0<\tau<\infty$, $p_i^\tau$ would further emphasizes more on the rare samples within each class by taking the local neighborhood structure into consideration. To shows this, we plot the $p_i^\tau$ curves for three samples from $\{circle, triangle \}$ classes in $\B$ by varying $\tau$ in Figure~\ref{fig:soft_assignment_weights}.
% %%%======
In the example shown in Figure~\ref{fig:soft_assignment_weights}, the blue and red circles come from the majority {circle} class, and the green triangle comes from the minority class. Often in typical sample weighting schemes, all samples within the same class are weighted uniformly and based on the inverse of their frequency. Thus samples in the minority class are always up-weighted compared to samples in the majority class. However, our sample weighting scheme allows for a more nuanced weighting. Compared with the blue circle, it can be observed that the red circle lies in the denser area, i.e. has more close neighbors. Hence, the $p_i^\tau$ of the red circle is consistently smaller than $p_i^\tau$ of the blue circle for the same $\tau \in (0, \infty)$. Also, they both converge to the inverse frequency $1/10$ when $\tau \rightarrow \infty$. When comparing samples from different classes, the green triangle (which is from the minority class) has more neighbors for smaller $\tau$, and thus could have a smaller weight compared to the blue circle which comes from the majority class. This allows for capturing a more nuanced notion of sample rarity within each class, which goes beyond the typical frequency-based methods.
Given $p_i^\tau$, the proposed CID method simply minimizing the objective~(\ref{eqn:obj}) such that $p_i^\tau$ is normalized using
$Z_{y_i}$ to equalize the total contribution of each class. Algorithm~\ref{alg:RBPN-O} describes the practical implementation of the proposed CID method in minimizing the objective~(\ref{eqn:obj}).
%This is vividly demonstrated from the red and blue $p_i^\tau$ curves in Figure~\ref{fig:soft_assignment_weights}. When it is the optimal $\tau$ that the neighborhood is a good approximate of the subpopulation of $g_i$, and the lower the density represents the minority protected groups. Hence $p_i^\tau$ emphasize on the worst protected groups. Then normalized by $Z_{y_i} = \sum_{i\in \B_{y_i}}p^{\tau}_{i}$ to guarantee the robust weights are equal for each class in the minimization objective~(\ref{eqn:obj}), the proposed CID method is also balance the robust weights between different classes to address the data imbalance problem.
% Next, we analyze the normalized variable $Z_{y_i}$ in the minimization objective, in which. By normalizing with $CZ_{y_i}$, the summation of the robust weight before the loss in ) equals to 1 over the whole batch $\B$ and equals to $1/C$ over each class sample $\B_y$. Having the same robust weights for each class can alleviate the class imbalanced issue introduced by the training task, while overall summation of robust weights equals to 1 is to ensure that the improvement does not come from the gradient scale shifting for fair comparison.
% \begin{table*}[htbp]
% \label{tab:id-highly-correlated-attributes}
% \caption{High ID correlated attributes. The blue attributes belong to the sensitivity groups.}
% \resizebox{\textwidth}{!}{
% \begin{tabular}{l|l} \hline
% CE & {\it [\textcolor{blue}{Male}, \textcolor{blue}{Young}, \textcolor{red}{Mouth Slightly Open}, Wearing Earrings, \textcolor{blue}{Oval Face}, Attractive, Eyeglasses, \textcolor{blue}{Pointy Nose}]}\\ \hline
% % IFW & {\it [\textcolor{blue}{Male}, Wearing Necktie, Sideburns, \textcolor{blue}{Big Nose}, Rosy Cheeks, Gray Hair, Wearing Necklace, Wearing Lipstick]} \\ \hline
% Sensitivity Attributes & {\it [Male, Pale Skin, Young, Oval Face, Big Lips, Big Nose, Pointy Nose, Narrow Eyes]} \\ \hline
% \end{tabular}
% }
% \end{table*}
\begin{figure}[t]
\centering
\includegraphics[width = \linewidth]{./figures/tau_illustrates.png}
\vspace{-0.2in}
\caption{Visualizing the effect of $\tau$ on $p_i^\tau$ and the soft local neighborhood in the proxy vector space $\{\z\}_{i=1}^{|\B|}$ and $\z\in \R^2$. We plot the $p_i^
\tau$ curves calculated according to equation~(\ref{eqn:p_ii}) by varying $\tau$ for the red and blue dots from the circle class, and the green sample from the triangle class. The circle class and triangle class include 10 samples and 4 samples, respectively.}
\label{fig:soft_assignment_weights}
\end{figure}
\vspace{-0.05in}
\section{Evaluation}
\vspace{-0.05in}
\label{sec:evaluation}
In addition to using standard classification accuracy metrics, we measure the robustness of the trained models using the following metrics.
\vspace{-0.05in}
\subsection{Area Under DEN Curves (AUD)}
\label{sec:AUDEN}
\vspace{-0.05in}
\cite{ardeshir2022estimating} introduces a metric for estimating structural performance disparity across an embedding space, referred to as \textit{Disparity across Embedding Neighborhoods} (DEN for short). The aforementioned study shows that such a metric is a good estimate of the performance disparity across groups when group information is not available. In fact, our objective function is specifically designed to minimize such disparity, thus we evaluate this metric on the test set to validate our assumption.
\vspace{-0.05in}
\subsection{Area Under Min-Max Curves (AUMM)}
\label{sec:AUIMMC}
\vspace{-0.05in}
As mentioned earlier, we do not have access to group labels during training, however, to measure if our model is in fact more robust across groups, we use group labels in the test set to validate our hypothesis. In our setup, we mostly focus on robustness/fairness across individuals, and given that the number of individuals in a face dataset could be very large, we define a modification to the widely used Rawlsian min-max metric. In the Rawlsian min max metric~\cite{rawls2001justice}, the ratio of the performance of the model is measured between the most and least accurate groups, i.e. $1-\frac{\min_{g}(e_g)}{\max_{g}(e_g)}|_{g\in G}$. This measure is often very useful when the number of groups is very limited. However, given that in our instance, we are interested in measuring disparity across different people, the number of different individuals in the dataset could be very large. Therefore, using the ratio of performance only on the highest and lowest individual will ignore large portions of the dataset. Thus we modify the Rawlsian min-max formulation to measure the ratio of the bottom-k\% and top-k\% of groups instead.
\vspace{-0.1in}
\begin{equation}
\text{MM} = \{1 - \frac{\bar{e}_g^k}{ \underline{e}_g^k}\}_{k=1}^{|G|}
\vspace{-0.1in}
\end{equation}
where $k\in [1,\cdots,|G|]$ denotes the index of groups. $\bar{e}_g^k$ and $\underline{e}_g^k$ are the average of top and bottom $k$ group performance, respectively. Sweeping k, results in a curve, which we refer to as the Min Max Curve. We use the area under this curve, AUMM for short, as a metric for robustness across groups. The lower the AUMM, the more robust/fair the model is.
%%%========
% %% should we move this to 5.2?
% In addition, we define an additional robustness metric, referred to as the area under the identity min-max curve (AUC-IMMC for short), which aims to measure such robustness in presence of a large number of categorical group labels. Two notions of disparity, standard deviation $\delta_G$ and the min-max ratio, RMM, of the best and lowest group performance have been widely
% used to evaluate group fairness. Hence, when the size of $|G|$ is getting large, for example $|G|$ equals to the total population on Earth for the identity attribute, and RMM becomes sensitive and invincible. To solve this problem, we propose a robust evaluation metric $\text{Min Max Curves (MMC)}$ that considers the ratio between the average performance of top $k$ groups and bottom $k$ groups, $\forall k \in [1,\cdots, |G|]$. Figure~\ref{fig:benefits_IMMC} demonstrates the advantage of MMC in quantifying the difference of four group accuracy distributions with the same RMM and $\delta_G$. The details will be discussed later in section~\ref{sec:IMMC}.
%%%=======
\begin{algorithm}[t]{\hspace*{-0.5in}}
\centering
\caption{CID Optimization ($\tau$)}
\label{alg:RBPN-O}
%\vspace{-0.1in}
\begin{algorithmic}[1]
\STATE Model initialization $\w_1$, proxy embeddings $\{\z_i\}_{i=1}^n$
\FOR {$t = 1,\ldots, T$}
\STATE Sample a batch of $B$ samples $\B=\{(\x_i,y_i)\}_{i}^{B}\sim\D$
\STATE Retrieve the proxy embedding vectors of batch samples, $\{\z_i\}_{i=1}^{B}$.
\STATE Calculate $p^\tau_{i}$ according to Eqn~(\ref{eqn:p_ii}) for $\forall (\x_i,y_i)\in \B$
\STATE Calculate $Z_{y_i} = \sum_{j\in \B_{y_i}} p^\tau_{j}$
\STATE Calculate CID loss: $\sum_{i\in\B}p^\tau_{i}\ell_i(\w_t)/(Z_{y_i})$
\STATE Update $\w_t$ using stochastic algorithms.
% \STATE When $\z$ is trainable:
% update $\z$.
\ENDFOR
\STATE {\bf Return} $\w_{T+1}$,
\end{algorithmic}
\end{algorithm}
%\subsection{Proxy Hard Mining Matters for Fairness }
%To let the model focus on the minority group. We doing the proxy hard mining in the fine-tunning stage such that we only optimize the samples with the largest robust weights. For fair comparison, in the baselines, we mining the samples with the largest losses.
%Mining the hardest losses is able to encourage the model emphasize more on the hardest samples that includes the underrepresented groups. However, the hard mining using proxy robust weights are able to encourage the model focus on the hard identities to boost the identity related fairness, such as gender. This explains why M-DRO achieves the best bottom 10\% testing identity accuracy.
%\begin{table*}[htbp]
%\centering
%\caption{Non-Conditional Accuracy Fairness evaluation metric of different methods. The results are averaged over 5 independent runs.The bold numbers denotes the best results among the baselines in each row.}
%\label{tab:celebA-pretraining}
%\begin{tabular}{l|c|c|c|c|c|c}
%\hline
%&CE &PRN+CE & DRO & PRN+DRO & M-DRO & M-PRN+DRO \\\hline
%Test ID Acc &- & -& -&-&0.9175 & \textbf{0.9177}\\
% Bottom 10 Percent ID Acc &0.8670 &0.8682&0.8678 & 0.8686 &\textbf{0.8689} & \underline{0.8688} \\
% ID Acc Std &0.02598& 0.02507&0.02578 &0.02575 &\textbf{0.0246}& \underline{0.02556}\\
% Gender RawMX&0.01866& 0.01821& 0.01849 &0.01834& 0.01843& \textbf{0.01736}\\
% Age RawMX &0.01985&0.01889&0.01920 & 0.01938 & 0.0190& \textbf{0.01881}\\ \hline
%\end{tabular}
%\end{table*}
\vspace{-0.05in}
\section{Experiments}
\vspace{-0.05in}
We evaluate the proposed approach alongside a few other baselines, on several datasets in terms of overall performance and robustness. In section \ref{sec:datasets} we go over the datasets and tasks used in our evaluation. In section \ref{sec:metrics} we provide details on our evaluation protocol and provide experimental results. Finally, in section \ref{sec:stress_test} we propose and report a stress test to measure robustness to controlled bias. In all experiments, for each face-image in the datasets, we extract its face-recognition embedding vector $\z$ using the face recognition model~\cite{facerecognitionmodel} and use it as its identity proxy.
\vspace{-0.05in}
\subsection{Datasets and Setup}
\vspace{-0.05in}
\label{sec:datasets}
We selected datasets that contained identity-independent tasks, and also contained information about the identity of the faces in the test set, in order to be able to evaluate the robustness of the trained models across identities at test time. As mentioned earlier, we do not use any identity information during the training phase, and solely rely on off-the-shelf face-recognition embeddings ~\cite{facerecognitionmodel} on the train set. In the following, we provide information on the three datasets used for experiments:\\
\textbf{CelebA}~\cite{liu2015faceattributes} has 200K face images and includes 10117 identities in total. Each image is labeled with 40 attributes/tasks. We pick two identity-independent tasks~\cite{liu2015faceattributes} \{{\it Smiling, Mouse Slightly Open(MSO)}\} and train standard binary classification models to predict those tests. We train ResNet18 model for 20 epochs both on SGD optimizers~\cite{bottou2012stochastic}. The learning rate is tuned in $\{0.003, 0.005, 0.01 \}$ for all the baselines. The hyperparameter $\tau$ in the CID method is tuned in $\{0.1:0.1:0.5 \}$.\\
\textbf{ExpW}~\cite{zhang2015learning} (cleaned version) is a facial expression dataset that includes 85K images with 1002 identities (split into 80\% train, 10\% val, and 10\% test), and 7 labels of facial expressions \{{\it angry, disgust, fear, happy, sad, surprise, neutral}\}.
To balance the size of the data scale and model capacity, we combine {\it sad-surprise-fear-neutral } as the new {\it ssfn} attribute to have enough positive samples ($\sim$ 19\%) to learn a valid CNN model. Then we predict \{{\it angry, disgust, happy, ssfn}\} expressions, respectively. Following the experimental setup in ~\cite{wang2019symmetric}, we adopt the SGD optimizer to optimize all a 4-layers CNN model. The structure of the model is provided in the appendix. We train 40 epochs using SGD optimizer. The learning rate is tuned $\{0.1, 0.05, 0.01\}$ and decayed at the 20th epoch by a factor of 100. $\tau$ is tuned in $\{0.1:0.1:0.5 \}$.
\\
\textbf{PugFig}~\cite{kumar2009attribute}\footnote{54K images of downloading URLs have been provided for the original dataset. But, most of them failed.} including 9K images with 111 identities (split into 80\% train, 10\% val, and 10\% test). Each image is tagged with the binary labels of lighting position, \{{\it frontal, non-frontal}\} and facial expression, \{{\it neutral, non-neutral}\}.
We train models to predict each task separately. Due to the limited data scale, we train the PubFig on the pre-trained ResNet18 model and fine-tune the fully connected layer for 60 epochs using SGD optimizer. The batch size is 16 and the weight-decay parameter is 5e-4. The learning rate is tuned in $\{0.005:0.001:0.01 \}$ and decayed at the 30th epoch by a factor of 10. $\tau$ is tuned in $\{0.1:0.1:0.5 \}$.\\
% \noindent
% \textbf{Data Split} The partitions of the dataset are provided in CelebA. We split the ExpW and PubFig across identities into train/val/test set with 80\%/10\%/10\% partitions.
% We show the same Id distribution on validation and test data in Figure~\ref{fig:identity_distribution}. Hence, experiments conducted on the default splits of CelebA are the In-Distribution (\textbf{ID}) settings.
% To show the effectiveness of CID on handling distribution shift, we construct Out-Of-Distribution (\textbf{OOD}) splits on CelebA by manipulating the artificial bias. The details for OOD CelebA subsets construction are described in Section~\ref{sec:stress_test}.
% For ExpW and PubFig dataset,
% the Ids are following heavy-tailed distributions on both datasets.
% We split the ExpW and PubFig into train/val/test set with 80\%/10\%/10\% partitions and in which train/val follows the heavy-tailed distributions and the test data follows the uniform distribution. To achieve this,
% we first random a certain number of identities $n$ and the sample each $k$ samples for each identity such that $nk = |\D|* 0.1$. Then we random split the rest of samples into the subset of train data with $|\D|* 0.9$ samples and val data with $|\D| * 0.1$. The distributions of Ids on train/val/test splits are described in Figure~\ref{fig:expw_pubfig}.
% \begin{figure}[h!]
% \centering
% % \includegraphics[width = 0.48\linewidth]{uniform_frequency_distribute.png}
% %\includegraphics[width = 0.48\linewidth]{uniform_IMMC.png}
% \includegraphics[width = 0.48\linewidth]{ExpW.png}
% \includegraphics[width = 0.48\linewidth]{./figures/pubfig.png}
% \caption{Ids distributions in the OOD splits of ExpW and PubFig}
% \label{fig:expw_pubfig}
% \end{figure}
% Tufts has 590 images with 118 identities in total and each identity has 5 images with 5 different expressions.
\vspace{-0.1in}
\subsection{Baselines}
\label{sec:baselines}
\vspace{-0.05in}
We compare proposed approach (CID) with the following effective baselines on fairness under awareness setup:
\textbf{IFW} (inverse frequency weighting)~\cite{huang2016learning, wang2017learning}: This is the typical sample weighting used to enforce class balance. Given $N_p, N_n$ number of positive and negative samples, the weights for the positive class and negative class samples are set proportional to $1/N_p$ and $1/N_n$, respectively.
\textbf{DRO} (Distributionally Robust Optimization)\cite{li2021tilted,qi2020attentional}: We implement the ABSGD~\cite{qi2020attentional} stochastic optimization method for optimizing $\max_\p\sum p_i\ell_i(\w)-\lambda\sum{p_i}\log np_i$ The hyperparameter $\lambda$ for DRO is tuned in $\{0.1, 0.5, 1, 2, 5\}$
\textbf{IRM} (Invariant Risk Minimization): we optimize $\ell(\w) + \lambda \|\nabla_{\upsilon|\upsilon=1.0} \ell(\upsilon\cdot \w)\|^2$ objective using the optimization framework of~\cite{arjovsky2019invariant}. $\lambda$ for IRM is tuned in $\{0.1:0.1:1\}$.
\textbf{ARL} (Adversarial Reweighted Learning)\cite{lahoti2020fairness}:
ARL optimizes $\min\limits_{\w}\max\limits_{\phi}p_i^{\phi} \ell_i(\w)$, where $p_i^\phi = 1/n +f_i^{\phi}/\sum_i f_i^{\phi}$, where $\phi$ is a linear adversary model with output score $f_i^\phi$\footnote{The details of implementation are provided in the supplimentary.}.
% y default, the sensitive groups are the identities of face images. In CelebA, it has XXXX identities in training while 1000 for validation and testing. Aff-Wild2 includes XXX in training and testing, respectively. ExpW XXXXX.
% Figure X describes distribution of the number of images per identity.
% \begin{table}[htbp]
% \centering
% \caption{Identity (Id) Partitions in Different Datasets}
% \label{tab:dataset_partition}
% \begin{tabular}{c|c|c|c}
% \toprule
% Id & Training & Validation & Testing \\ \hline
% CelebA & 8192 & 985 & 1000\\
% ExpW & 1002 & 500 & 500 \\
% PubFig & 140 & 140 & 140 \\
% \bottomrule
% \end{tabular}
% \end{table}
\vspace{-0.05in}
\subsection{Measuring Robustness}
\vspace{-0.05in}
\label{sec:metrics}
We evaluate the performance of the baselines mentioned in Section \ref{sec:baselines}, in terms of overall classification accuracy (Acc), average, and standard-deviation of per-identity accuracy (Id Acc and $\delta_{\text{Id}}$). This is evaluated by measuring a model's prediction accuracy for each identity (person) in the test set, and reporting mean and standard deviation. Ideally a high-performing and identity-robust model should maintain high Id Acc, while having low $\delta_{\text{Id}}$. A low $\delta_{\text{Id}}$ is one of the metrics implying that the performance of the model is robust across identities and thus more fair. In addition, we report the accuracy on the least accurate 10\% of identities in the test set. Plus, we evaluate the model in terms of area under DEN curve (referred to as AUD for short), proposed in \cite{ardeshir2022estimating}, and explained in section \ref{sec:AUDEN}, which measures the disparity of performance across different neighborhoods of the embedding space of the face-recognition embedding. The lower the AUD, the more robust the model is. Also, as described in section \ref{sec:AUIMMC}, we use the area under the min-max curve (AUMM for short) as another robustness metric. Given that this metric measures the disparity between the accuracy of the top-k and bottom-k identities, the lower the AUMM, the more identity-robust a model is.
Table~\ref{tab:ID_CelebA}, \ref{tab:expW}, and \ref{tab:pubfig} report experimental results on CelebA, ExpW, and PubFig datasets respectively, and averaged over 5 independent runs. It can be observed that CID outperforms all the baselines in terms of every model robustness metric, namely, (lowest) 10\% Id Acc, $\delta_{\text{Id}}$, AUMM, and AUD. In addition, the overall accuracy is also either higher than baseline, or very competitive. In other words, CID does maintain high accuracy while attaining robustness.
\begin{figure}[htbp]
\centering
\includegraphics[width = 0.45\linewidth]{./figures/SME_k.jpg}
\includegraphics[width = 0.45\linewidth]{./figures/MSO_k.jpg}
\caption{The Min-Max curves for the {\it Smiling} and {\it Mouth Slightly Open (MSO)} tasks. It can be observed that CID consistently yields a lower curve, resulting in a smaller area under the MM curve, and thus, less disparity across top and bottom k\% identities.}
\label{fig:IMMC_Default_Data}
\end{figure}
\begin{table}[h!]
\caption{Attributes Prediction Experimental Results on CelebA. {\bf Bold} and \underline{underline} are best and second-best results of each metric.}
\label{tab:ID_CelebA}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Smiling} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & \textbf{92.73} & \underline{92.02} & 71.02 & 0.0962 & \underline{0.1544} & 0.1156 \\
IFW & 92.69 & \textbf{92.99} & 71.03 & 0.0968 & 0.1550 & \underline{0.1141} \\
DRO & 92.55 & 91.90 & 71.06 & 0.0965 & 0.1546 & 0.1146 \\
IRM & 92.71 & 92.11 & \underline{ 71.74 } & 0.0939 & 0.1522 & 0.1143 \\
ARL & 92.71 & 92.07 & 70.94 & 0.0975 & 0.1558 & 0.1151 \\ \hline
CID &\underline{92.72} & \underline{92.15} & \textbf{71.94} & \textbf{0.0926} & \textbf{0.1506} & \textbf{0.1126}\\
% CID+IFW & 92.69 & 91.98 & \underline{71.05} & \underline{0.09614} & 0.1551 \\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf MSO } & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 94.09 & 93.71 & 74.57 & 0.0838 & 0.1322 &0.1036 \\
IFW & \underline{94.09} & \underline{93.73} & 75.06 & \underline{0.0832} & \underline{0.1305} & \underline{0.1019} \\
DRO & 94.04 & 93.69 & 75.07 & 0.0840 & 0.1341 &0.1038 \\
IRM & 94.04 & 92.68 & \underline{75.10} & 0.0837 & 0.1325 & 0.1021 \\
ARL & 93.97 & 93.54 & 74.04 & 0.0871 & 0.1376 & 0.1033 \\
\hline
% CID & \textbf{94.14} & \underline{93.75} & \underline{75.23} & 0.0838 & 0.1308 \\
% CID+IFW & \underline{94.13} & \textbf{93.79} & \textbf{75.36} & \textbf{0.0824} & \textbf{0.1299} \\
CID & {\bf 94.13} & \textbf{93.79} & \textbf{75.36} & \textbf{0.0824} & \textbf{0.1299} & \textbf{0.1005 } \\
\bottomrule
\end{tabular}
}
\end{table}
\begin{table}[h!]
\caption{Expression Prediction Experimental Results on ExpW. }
\label{tab:expW}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Angry} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & \underline{63.25} & \underline{ 63.26} & \underline{ 33.46} & 0.1706 & 0.3603 & 0.2270 \\
IFW & 60.83 & 60.84 & 32.29 & \underline{0.1691} & 0.3643 & 0.2309\\
DRO & 63.18 & 63.19 & 33.26 & 0.1693 & 0.3622 & \underline{0.2205} \\
IRM & 63.17 & 63.18 & 33.20 & 0.1712 & 0.3615 & 0.2232 \\
ARL & 63.44 & 63.45 & 33.43 & 0.1693 & \underline{0.3584} & 0.2214 \\\hline
CID & \textbf{ 63.53} & \textbf{63.53} & \textbf{33.53 } & \textbf{0.1672} & \textbf{0.3543}& \textbf{0.2188}\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Disgust} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 75.23 & 75.23 & \underline{ 46.19} & 0.1517 & \underline{0.2862} & 0.1875 \\
IFW & 74.93 & 74.94 & 44.99 & \underline{0.1514} & 0.2938 & \underline{0.1758}\\
DRO & \underline{75.26} & \underline{ 75.27} & 46.09 & 0.1520 & 0.2865 & 0.1867 \\
IRM & 75.25 & 75.27 & 46.00 & 0.1518 & 0.2863 & 0.1921\\
ARL & 75.13 & 75.14 & 45.80 & 0.1531 & 0.2893 & 0.1911\\\hline
CID & \textbf{75.27} & \textbf{75.28} & \textbf{46.22} & \textbf{0.1503} & \textbf{0.2837} & \textbf{0.1749} \\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Happy} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
CE & 86.94 & 86.95 & 59.28 & 0.1292 & 0.2159 & 0.1429
\\
IFW & 86.86 & 86.87 & 59.24 & 0.1290 & 0.2160 & 0.1326 \\
DRO & 86.95 & 86.96 & 59.34 & \underline{0.1286} & 0.2151 & \underline{0.1311} \\
IRM & \underline{86.98} & \underline{87.00} & \underline{59.73} & 0.1289 & \underline{0.2148} & 0.1386\\
ARL & 86.73 & 86.95 & 0.5957 & 0.1290 & 0.2166 & 0.1391\\ \hline
CID & \textbf{87.00} & \textbf{87.01} & \textbf{59.99} & \textbf{0.1281} & \textbf{0.2140} & \textbf{0.1251} \\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf SSFN} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & \underline{85.16} & \underline{85.17} & 58.85 & 0.1273 & \underline{ 0.2187} & 0.1583\\
IFW & 79.98 & 79.99 & 53.13 & 0.1434 & 0.2588 & 0.1695 \\
DRO & 85.19 & 85.20 & 58.95 & 0.1278 & 0.2200 & \underline{0.1583}\\
IRM & \textbf{85.20} & \textbf{ 85.21} & \underline{59.08} & \underline{0.1271} & 0.2197 & 0.1632 \\
ARL & 85.07 & 84.51 & 57.52 & 0.1342 & 0.2259 & 0.1655 \\\hline
CID & 85.14 & 85.15 & \textbf{59.60} & \textbf{0.1260} & \textbf{0.2163} & \textbf{0.1576}\\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}[h!]
\caption{Experimental Results on PubFig. }
\label{tab:pubfig}
\centering
%\vspace{-0.1in}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Frontal} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 74.83 & 74.93 & 36.11 & \underline{0.1913} & 0.3487 & \underline{0.1639} \\
IFW & 71.68 & 71.77 & 34.72 & 0.1969 & 0.3582 & 0.1837 \\
DRO & 74.60 & 74.70 & 37.15 & 0.1930 & 0.3491 & 0.1660\\
IRM & \underline{75.13} & \underline{75.24} & \underline{37.85} & 0.1974 & \underline{0.3465} & 0.1657 \\
ARL & 73.56 & 74.77 & 37.85 & 0.1931 & 0.3479 & 0.1655\\ \hline
CID & \textbf{75.10} & \textbf{75.20} & \textbf{ 38.19} & \textbf{ 0.1906 } & \textbf{ 0.3449} & \textbf{0.1606}\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf Neutral} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 56.48 & 56.55 & 26.74 & 0.1798 & \underline{0.3985} & 0.2645 \\
IFW & \underline{56.61} & \textbf{56.69} & 26.74 & \underline{0.1773} & 0.4050 & 0.2593 \\
DRO & 56.24 & 56.32 & 26.74 & 0.1782 & 0.4034 & 0.2588 \\
IRM & 56.48 & 56.55 & \underline{27.08} & 0.1817 & 0.4009 & \underline{0.2582}\\
%\textcolor{blue}{ARL} & 58.52 & 53.60 & 27.08 & 0.1690 & 0.3910 & -\\
ARL & \textbf{56.91} & 55.85 & 26.04 & 0.1772 & 0.3995 & 0.2575\\
\hline
%CID & \underline{56.51} & \underline{56.59} & \textbf{ 28.82} & \textbf{0.1748} & \textbf{0.3880} & \textbf{0.2544} \\
CID & 56.51 & \underline{56.59} & \textbf{ 27.43} & \textbf{0.1677} & \textbf{0.3864} & \textbf{0.2544} \\
\bottomrule
\end{tabular}
}
\end{table}
%% Saved Sentences for DEN Curves
% According to section~\ref{sec:IMMC}, AUM requires known ground truth identity labels in the test dataset. In most cases, however, current face-related datasets are collected for the general computer vision tasks, such as attribute prediction and segmentation tasks, rather than the fairness-specific tasks in the first place. The fairness-related sensitive group label information may not even be available for validation.
% The DEN curves have been proposed in~\cite{ardeshir2022estimating} to estimate the structural proxy neighbourhood disparity. When the ground truth sensitivity labels are even unknown for evaluation, we report the area under the DEN curves.
% Therefore, instead of AUM, we report the approximated structural DEN disparity from~\cite{ardeshir2022estimating} when the ground truth identity labels are not available for evaluation.
\vspace{-0.1in}
\subsection{Stress-testing with Controlled Bias}
\vspace{-0.05in}
\label{sec:stress_test}
In this section, we aim to measure how sensitive to dataset bias a model is by stress testing the model. To do so, we construct different versions of the train/validation sets of CelebA, by manipulating the dataset and adding controlled artificial identity-to-task bias. More specifically, given a task (such as smiling), we construct a biased train set by excluding $p$\% of data points belonging to a (task-label, subpopulation). As an example, in one of the variations, we exclude 50\% of male-smiling images. This would artificially create a biased dataset that is prone to correlating male faces to the label non-smiling. If we train a classifier on such a dataset, a model's performance on the standard (non-manipulated) test set can be very non-robust, as the train and test set do not follow the same distribution. In all the setups we solely manipulate the bias in the train set and keep the test set unchanged. We try this experiment for different values of $p \in \{25\%, 50\%, 75\%, 90\%\}$, and for different (group, task-label) combinations. We refer to each setting by specifying which group (M:Male vs. F:Female), and which task label (P: positive, N: negative) has been manipulated (excluded by p\%) from the training and validation set (while the test set is unchanged). As an example, on the task "Smiling", FP 50\% means that half of the Female Positives (smiling female faces) were excluded during training, therefore creating a bias in the dataset. Combinations of these setups could also be generated to further exaggerate the bias (such as FPMN: removing positive/smiling female images and negative/non-smiling male images). Figure~\ref{fig:OOD_results} shows the results of our stress test on the task "Smiling". Due to space limitations, we only report the robustness measures, namely, MMC, $\delta_{\text{Id}}$, AUMM, and bottom 10\% Id Acc experimental results between CE (cross-entropy) and CID on all different setups and levels of induced bias. We also provide more stress test results, including other metrics, analysis on the other CelebA task, and other variations of the manipulation setups in the supplementary. All other variations of tasks, metrics and setups follow the same trend.
% \begin{figure*}[htbp]
% \centering
% \includegraphics[width = 0.18\linewidth]{./figures/val_id_0_dis_celeba.png}
% \includegraphics[width = 0.18\linewidth]{./figures/val_id_25_dis_celeba.png}
% \includegraphics[width = 0.18\linewidth]{./figures/val_id_50_dis_celeba.png}
% \includegraphics[width = 0.18\linewidth]{./figures/val_id_75_dis_celeba.png}
% \includegraphics[width = 0.18\linewidth]{./figures/val_id_90_dis_celeba.png}
% \caption{Identity frequency distribution shifting between ID and OOD of CelebA. DFS 25\% represents the OOD distribution that deleting 25\% percentage of Female Smiling images from the CelebA.}
% \label{fig:freq_id_distribution}
% \end{figure*}
\begin{table}[htbp]
\caption{Stress testing the models on the CelebA dataset, by eliminating 90\% of a subpopulation in the training/validation set.}
\label{tab:OOD_CelebA}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf FP} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 90.33 & 89.77 & 65.91 & 0.1095 & 0.1821 &0.1354 \\
IFW & 91.13 & \underline{90.49} &\underline{67.28} & \underline{0.1053} & \underline{0.1737} & \underline{0.1259}\\
DRO & 90.23 & 89.84 & 66.20 & 0.1098 & 0.1834 & 0.1367\\
IRM & \underline{91.18} & 90.11 & 66.54 & 0.1088 & 0.1789 & 0.1343 \\
ARL & 90.40 & 89.46 & 65.55 & 0.1127 & 0.1863 & 0.1323 \\ \hline
CID & \textbf{91.31} & \textbf{90.67} & \textbf{67.73} & \textbf{0.1042} & \textbf{0.1717} & \textbf{0.1233}\\
% CID+IFW & \underline{91.27} & \textbf{90.67} & \textbf{67.76} & \textbf{ 0.1030} & \textbf{0.1717} & 0.1295\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf FN} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 89.42 & 89.51 & 66.52 & 0.1077 & 0.1821 & 0.1476 \\
IFW & \underline{90.47} & \underline{ 90.29} & \underline{67.45} & \underline{0.1074} & \underline{0.1761} & \underline{0.1281} \\
DRO & 89.43 & 89.51 & 66.53 & 0.1077 & 0.1817 & 0.1501 \\
IRM & 90.18 & 89.91 & 66.84 & 0.1075 & 0.1799 & 0.1438 \\
ARL & 89.64 & 89.66 & 66.78 & 0.1075 & 0.1807 & 0.1456 \\ \hline
CID & \textbf{90.89} & \textbf{90.58} & \textbf{68.12} & \textbf{0.1057} & \textbf{0.1727} & \textbf{0.1241} \\
% CID+IFW & \textbf{91.59} & \textbf{91.03} & \textbf{69.22} & \textbf{0.1014} & \textbf{0.1658} & \textbf{0.1228}\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf MP } & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
CE & \underline{91.51} & 90.03 & 64.50 & 0.1162 & 0.1885 & 0.1349 \\
IFW & 91.34 & \underline{90.35} & \underline{65.69} & \underline{0.1127} & \underline{ 0.1820} & \underline{0.1309} \\
DRO & 91.12 & 90.13 & 64.70 & 0.1172 & 0.1855 & 0.1381 \\
IRM & 91.22 & 90.23 & 65.01 & 0.1168 & 0.1823 & 0.1361 \\
ARL & 91.27 & 90.17 & 64.49 & 0.1169 & 0.1890 & 0.1371 \\ \hline
CID & \textbf{91.58} & \textbf{90.62} & \textbf{66.68} & \textbf{0.1091} & \textbf{0.1774} & \textbf{0.1221} \\
% CID+IFW & \textbf{91.58} & \textbf{90.85} & \textbf{67.56} & \textbf{0.1083} & \textbf{0.1742}& \textbf{0.1219}\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf MN} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
CE & 90.29 & 89.53 & 63.72 & 0.1151 & 0.1922 & 0.1433 \\
IFW & 90.07 & \underline{90.24} & \underline{65.68} & \underline{0.1093} & \underline{0.1816} & \underline{0.1372} \\
DRO & 90.03 & 89.60 & 63.91 & 0.1146 & 0.1915 & 0.1490 \\
IRM & 90.06 & 89.70 & 64.11 & 0.1123 & 0.1845 & 0.1455\\
ARL & \underline{90.33} &89.51 & 63.54 & 0.1162 & 0.1956 & - \\ \hline
CID & \textbf{91.20} & \textbf{90.43} & \textbf{66.16} & \textbf{0.1080} & \textbf{0.1783} & \textbf{0.1241}\\
% CID+IFW & \textbf{91.29} & \textbf{90.45} & \textbf{66.57} & \textbf{0.1070} & \textbf{0.1768} & \underline{0.1249}\\
\bottomrule
\end{tabular}}
\resizebox{\linewidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\toprule
{\bf FPMN } & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
CE & 87.25 & 86.48 & 57.28 & 0.1295 & 0.2229 & \underline{0.1584} \\
IFW & 87.23 & 86.41 & 57.66 & 0.1322 & 0.2279 & 0.1655 \\
DRO & \underline{87.37} & \underline{86.62} &
\underline{59.14} & \underline{0.1280} & \underline{0.2208} & 0.1625 \\
IRM & 87.36 & 86.63 & 59.13 & 0.1283 & 0.2210 & 0.1614 \\
ARL & 87.32 & 86.55 & 58.33 & 0.1303 & 0.2247 & 0.1631 \\ \hline
CID & \textbf{87.87} & \textbf{87.09} & \textbf{60.24} & \textbf{ 0.1264} & \textbf{0.2170} &\textbf{0.1553}\\
% CID+IFW & \underline{87.73} & \underline{86.87} & 58.17 & 0.1311 & 0.2241 & 0.1665 \\
\bottomrule
\end{tabular}}
% \resizebox{\linewidth}{!}{
% \begin{tabular}{c|c|c|c|c|c|c}
% \toprule
% {\bf FNMP } & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
% CE & 87.61 & 87.28 & 60.47 & 0.1252 & 0.2137 & \underline{0.1479 }\\
% IFW & 87.75 & 86.28 & 60.13 & 0.1256 & 0.2161 & 0.1533 \\
% DRO & \underline{87.85} & \underline{87.45} &
% \underline{60.64} & \underline{0.1240} & \underline{0.2122} & 0.1535 \\
% IRM & 87.77 & 86.23 & 60.29 & 0.1256 & 0.2145 & 0.1493 \\ \hline
% CID & \textbf{88.44} & \textbf{87.93} & \textbf{61.54} & \textbf{0.1224} & \textbf{0.2068} &\textbf{0.1465}\\
% \bottomrule
% \end{tabular}}
\end{table}
\begin{table}[h!]
\centering
\caption{Accuracy of the model on the least accurate subpopulation in the CelebA dataset \{{\it task-label$\times$Male/Female}\} on the stress testing CelebA dataset. It can be observed that CID achieves the highest worst-case accuracy, and therefore the most robust among the evaluated baselines.}
\label{tbl:worst_group_domain_shifts}
\resizebox{0.7\linewidth}{!}{
\begin{tabular}{c|c|c|c|c|c}
\toprule
& FP & FN & MP & MN & FPMN \\ \hline
CE & 83.32 & 76.96 & 72.87 & 78.09 & \underline{76.24} \\
IFW &\underline{87.87} &\underline{80.24} &\underline{ 75.46} & \underline{82.46} & 73.34 \\
DRO & 83.16 & 76.98 & 73.05 & 78.40 & 76.19 \\
IRM & 83.77 & 77.04 & 73.51 & 78.33 & 75.93 \\
ARL & 83.09 & 76.47 & 74.37 &78.32 & 75.62 \\\hline
CID & \textbf{88.26} & \textbf{82.95} & \textbf{77.51} & \textbf{83.88} & \textbf{76.47} \\
\bottomrule
\end{tabular}}
\vspace{-0.1in}
\end{table}
\begin{figure*}[t!]
\centering
\includegraphics[width = 0.21\linewidth]{./figures/DFS.jpg}
\includegraphics[width = 0.18\linewidth]{./figures/DFN.jpg}
\includegraphics[width = 0.18\linewidth]{./figures/DMS.jpg}
\includegraphics[width = 0.18\linewidth]{./figures/DMN.jpg}
\includegraphics[width = 0.18\linewidth]{./figures/DFSMN.jpg}
\includegraphics[width = 0.22\linewidth]{./figures/DFS_IMMC_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFN_IMMC_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMS_IMMC_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMN_IMMC_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFSMN_IMMC_share_yaxis_biased_ptg.png}
% \caption{The area under identity rawlsian min-max curves (AUC IMMC) on different biased group distributions.}
% \label{fig:AUC_IMMC_biased_ptg_trend}
% \end{figure*}
% \begin{figure*}[h]
% \centering
\includegraphics[width = 0.225\linewidth]{./figures/DFS_STD_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFN_STD_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMS_STD_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMN_STD_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFSMN_STD_share_yaxis_biased_ptg.png}
% \caption{The standard deviation of ID accuracy on different biased group distributions.}
% \label{fig:STD_biased_ptg_trend}
% \end{figure*}
% \begin{figure*}[h]
% \centering
\includegraphics[width = 0.22\linewidth]{./figures/DFS_Bottom10_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFN_Bottom10_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMS_Bottom10_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DMN_Bottom10_share_yaxis_biased_ptg.png}
\includegraphics[width = 0.18\linewidth]{./figures/DFSMN_Bottom10_share_yaxis_biased_ptg.png}
\caption{Results of different stress tests on the CelebA dataset for the smiling task label. The identity MM curves, area under identity min-max curves (AUMM), standard deviation of Id accuracy $\delta_{\text{Id}}$, and the bottom 10\% percentage Id accuracy are measured for different setups and under different levels of induced bias. First row: For the MM figure, the x-axis shows k for which the disparity between top and bottom k identities is evaluated. the In each figure, x-axis specifies the amount (percentage) of the training data of the (group, task-label) that is excluded during training. As it can be observed in all instances. As it can be observed, CID maintains its original metrics significantly better than CE in the presence of distribution shift.}
\label{fig:OOD_results}
\end{figure*}
We observe that training models on these biased versions would cause their test performance to reduce (degrade) as the amount of bias increases. Similarly, disparity metrics increase (degrade) with more bias, as expected. This shows that our stress-testing framework in fact does introduce controlled bias to the trained model. We make the following observations from these experiments: 1)
For each \{{\it task-label $\times$ Male/Female}\} combination, the higher the amount of manipulation, the higher the MMC and AUMM values, which implies that the proposed metrics (MMC and AUMM) do capture the amount of bias in a model. 2) The proposed CID method has smaller MMC, AUMM, and $\delta_{\text{Id}}$ values, in addition to higher bottom 10\% Id accuracy compared to CE. The gap between the models steadily increases as the amount of induced bias in the dataset increases, which verifies the advantages of CID on handling distribution shifting over Empirical Risk Minimization.
To narrow down the scope, we report experimental results that compare CID with more baseline methods on the most biased version of each stress-test setup (i.e, MP 90\%, MN 90\%, FP 90\%, FN 90\%, and FPMN 90\% in Table~\ref{tab:OOD_CelebA}). Please note, given that we need access to group labels such as Male/Female and identity labels, we were only able to conduct this specific experiment on the CelebA dataset. To conclude this stress-test, all robustness metrics are significantly more desirable for CID compared to the baselines, alluding to the fact that our CID weighting scheme successfully mitigates bias, and maintains high accuracy in presence of it. To make our experiments complement, we report the worst group accuracy that is widely used in the baseline comparison in terms of \{{\it task-label $\times$ Male/Female}\} in Table~\ref{tbl:worst_group_domain_shifts}.
% \subsection{Observations}
% \label{sec:celeba}
% \noindent
% % \textbf{Setting}
% % We train ResNet18 model for 20 epochs both on ID and OOD CelebA with SGD optimizers~\cite{bottou2012stochastic}. The learning rate is tuned in $\{0.003, 0.005, 0.01 \}$ for all the baselines. The hyperparameter $\tau$ in the CID method is tuned in $\{0.1:0.1:0.5 \}$.
% % \noindent
% % \textbf{Experimental Results}
% Table~\ref{tab:ID_CelebA}, \ref{tab:expW}, and \ref{tab:pubfig} report the experimental results on celebA, expW, and PubFig datasets respectively, and averaged over 5 independent runs. It can be observed that CID outperforms all the baselines in terms of all model robustness metrics, namely, (lowest) 10\% Id Acc, $\delta_{\text{Id}}$, AUMM, and AUD. In addition, the overall accuracy is also either higher than baseline, or within the margin of error. In other words, CID does maintain high accuracy while attaining ronustness.
% Table \ref{tab:OOD_CelebA} contains detailed results on the stress test performed on the celebA dataset and described in Section \ref{sec:stress_test}, which shows similar trends but with higher margins.
% When it comes to OOD subsets of CelebA settings reported in Table \ref{tab:OOD_CelebA}, we first claim that besides the identity distribution shift, FP, FN, MP, and MN subsets
% of CelebA also exaggerate the class imbalance by either deleting Smiling or Non-smiling images. Hence, the classed weighted IFW method performs best among the baselines in FP, FN, MP and MN implies that addressing the class imbalance distribution shift would improve model robustness to some extent when the group truth groups $G$ are not available. Furthermore, the method that combines CID and IFW, i.e, CID+IFW, outperforms CID consistently in those settings verifies the effectiveness of improve model robustness by alleviating the distribution shifts both in terms of identities and task class labels at the same time.
% In the DFSMN settings, CID achieves the best performance as DFSMN only exaggerate the identity distribution shifts as both Smiling and Non-smiling images are deleting at the same time.
% \subsection{Experimental Results on ExpW}
% \noindent
% \textbf{Settings}
% % In ExpW dataset, we adopt the SGD optimizer to optimize all a 4-layers CNN model. The structure of the model is provided in the appendix. We train 40 epochs using SGD optimizer. The learning rate is tuned $\{0.1, 0.05, 0.01\}$ and decayed at the 20th epoch by a factor of 100.
% \noindent
% \textbf{Experimental Results}
% As we have directly imposed the domain shifts across train/val and test splits (as explained in XXXXX), we observe the outstanding performance of CID over the other baselines shows the effectiveness of CID in ...
% \subsection{Experimental Results on PubFig}
% % \noindent\textbf{Settings} Due to the limited data scale, we train the PubFig on the pretrained ResNet18 model and fine-tune the fully connected layer for 60 epochs with SGD optimizer. The batch size is 16 and the weight-decay parameter is 5e-4. The learning rate is tuned in $\{0.005:0.001:0.01 \}$ and decayed at the 30th epoch by a factor of 10.
% \noindent
% \textbf{Experimental Results}
\vspace{-0.1in}
\section{Conclusion}
\label{sec:conclusion}
\vspace{-0.1in}
We propose a framework to effectively use off-the-shelf face-recognition model embeddings to improve the robustness/fairness of identity-independent face models. Our experiments show that our simple sample-weighting approach helps face models to maintain high accuracy while gaining significant robustness to distribution shifts and different levels of bias, and often maintaining a more uniform performance across different identities (and groups) of faces.
% \subsection{Attribute Statistics}
% In Table~X, we report the statistics of each group in terms of
% Table~\ref{tab:attribute-statistics} are the statistics of each group in terms $\{Smiling \ \times \ Gender\}$. And we report the validation and testing data statistics in the appendix. By comparing different group statistics, train, validation and test dataset are consistent in terms of different groups, which belongs to \textbf{ID} experimental setting.
% Therefore, to measure the performance of our method in the \textbf{OOD} setting that wildly existed in real world applications, we enlarge the distribution gap between training and test data by removing different percentages of certain group samples in the training data and validation data, which the test data remains unchanged.
% We delete $25\%$, $50\%$, $75\%$, $90\%$ percentage of samples from three different group settings Female Smiling (FS), Male Non-smiling (MN), and Female Smiling & Male Non-smiling (FSMN), respectively. For example, DFSMN 50\% represents that we delete 50\% of images both from the Male Non-smiling group samples and Male Non-smiling group samples in the train and validation dataset. DFS 50\% and DMN 50\% follow the same naming rules. The higher the percentage, the larger distribution drifting between training/validation data and test data.
%The same as \cite{liu2021just}, the {\it worst group} is defined as the lowest testing accuracy group trained by standard CE loss. We observed that {\it the worst group} is inverse proportion to the number of samples.
% \begin{table}[h!]
% \centering
% \caption{Smiling mutable feature attribute statistics}
% \label{tab:smile-statistics}
% \begin{tabular}{c|cc|c}
% \hline
% & Smiling & Non-smiling & Gender \\ \hline
% Female & 50488 & 42175 & 92663 \\
% Male & 26921 & 38564 & 65485 \\ \hline
% Smile & 77409 & 80739 & 158148 \\
% \hline
% \end{tabular}
% \end{table}
% In addition, we decompose the improvements of the CID from the class level balanced normalization reweighting and the fairness structure information provided by face proxy in $p_i^{\tau}$. To achieve this, we conduct \textbf{BAC} setting experiments by deleting part of samples from FS to have a equal number of samples with MN, while delete part of the samples from MN to have equal number samples with MS, which leads to the balanced statistics in Table~\ref{tab:balanced-smile-statistics}. In this way we are able to measure the fairness improvements induced by face proxy.
% In addition, we either report Male or Female attribute accuracy in different settings depending on the number of samples in the training.
% \subsection{Deletes Half of Female Smiling and Half Male Non-Smiling}
% We delete half of samples from female smiling group and non-smiling group. According to the cross entropy results, {\it \{MS $\leq$ FS, MN\} $<<$ FN} in the screwed testing data set.
% \begin{table}[htbp]
% \centering
% \caption{DFSMN}
% \begin{tabular}{c|cccc}
% \hline & CE & CID & IFW & CID+IFW \\ \hline
% Mean Acc & 0.9140 & 0.9144 & \underline{0.9148} & 0.9140 \\
% Mean ID Acc & 0.9067 & 0.9055 &\underline{0.9075} & 0.906 \\
% 10\% ID Acc & 0.6736 & 0.6667 & 0.6870 & \underline{0.6787} \\
% ID Acc Std & 0.1058 & 0.1103 & 0.1053 & \underline{0.1028} \\
% Worst Group Acc & & & & \\
% F-BAcc & & & & \\
% M-BAcc & & & & \\
% Best Epoch & 17 & 18 & 18 & 18 \\ \hline
% \end{tabular}
% \end{table}
% \newpage
% \begin{table}[h!]
% \centering
% \caption{Balanced smiling mutable feature attribute statistics}
% \label{tab:balanced-smile-statistics}
% \begin{tabular}{c|cc|c}
% \hline
% & Smiling & Non-smiling & Gender \\ \hline
% Female & 42175 & 42175 & 84350 \\
% Male & 26921 & 26921 & 53842 \\ \hline
% Smile & 69069 & 69069 & 138192 \\
% \hline
% \end{tabular}
% \end{table}
% \begin{table*}[htbp]
% \centering
% \caption{Experimental results.~\textcolor{red}{ IRM, DRO, CID+ IFW results}}
% % Random Seed: 1, 3, 6
% %lr: 0.001
% %batch size: 256
% \begin{tabular}{c|cc|cc|cc}
% \toprule
% & Smile CE & Simle CID & MSO CE & MOS CID & EGS CE & EGS CID \\ \hline
% Acc & 0.9270 &\textbf{ 0.9279 } &\textbf{0.9413} &0.9410 & 0.9968 & \textbf{0.9966} \\
% Id Acc & 0.9198 & \textbf{0.9214} &0.9363 &\textbf{0.9367} & 0.9953 & \textbf{0.9956} \\
% 10\% Id Acc & 0.7077 & \textbf{0.7176} &0.7334 & \textbf{0.7416} & 0.9724 &\textbf{0.9759} \\
% %Minimum Group Accuracy & 0.8925 &\textbf{ 0.8964} & 0.9331& \textbf{ 0.9358}&0.9766& \textbf{0.9785}\\
% %CE Minimum Group Accuracy & 0.9185 & \textbf{0.9187} & \underline{0.9443}& \underline{0.9443}& 0.9936 & 0.9936 \\ \hline
% $\delta_{\text{Id}}$ & 0.0972 & \textbf{0.0935} & 0.0903& \textbf{0.0865}& 0.0311 & \textbf{0.0279} \\
% AUIMC & 0.1556 & \textbf{0.1506} &0.1384 & \textbf{0.1360} &0.0235 & \textbf{0.0216} \\
% AUDEN &0.6159 &\textbf{0.6139} &0.6413 &\textbf{0.6354} & 0.3413 &\textbf{0.3354} \\ \bottomrule
% \end{tabular}
% \end{table*}
%\Floatbarrier
\newpage
% \section{Experiments}
% \subsection{Settings}
% We adopt two stage training method. In the first stage, we training the ResNet18 model using CE until converge. To improve the classification head capacity, we add one more MLP with a following ReLu activation layer before the classifier in the finetuning stage. Then we finetunes the last block of the model, and classifiers for ResNet18.
% \subsection{Baselines}
% Here we consider two baselines, CE and DRO. The CE are standard batch training scheme for images and DRO [XXX] are designed to encourage the models focus on the hard samples by assigning larger weights to the larger losses. Our PRN is flexible to adptive to both methods, %denoted as PRN+CE, and PRN+DRO. For the hard mining methods, we only applying on the top of DRO and NEB+PRN methods, which we denote as M-DRO and M-PRN+DRO, respectively. For all methods, w do not use any information about gender and age. The ground truth labels are only used for evaluation.
% Then we report the overall identity accuracy, bottom 10\% identity accuracy, standard deviation of per identity accuracy, gender Rawsian Min Max and age Rawsian Min Max.
% \subsection{Results}
% Table~\ref{tab:celebA-pretraining} reports the results on the testing dataset. For the CE and DRO methods, it is obvious to see that the added PRN method improves its corresponding methods for all metrics. The M-PRN+DRO has the best results in terms of test identity accuracy, gender rawlsian min-max, and age rawlsian min-max, while the M-DRO achieves the best bottom 10 percent identity accuracy. Compared with M-PRN+DRO, the improved bottom identity accuracy and lower overall testing identity accuracy automatically leading to a small standard deviation of identity accuracy for M-DRO method.
% Then by comparing with standard CE and DRO methods, M-PRN+DRO has relative improvements of 6.9\% and 4.5\% on gender Rawlsian min-max, and 5.2\% and 2.0\% on age Rawlsian min-maxïŒ respectively. In addition, the face proxy has a clear decision boundary in terms of gender attribute while noisy decision boundary in terms of age attribute explains the relatively higher improvements on gender than age attribute.
{\small
\bibliographystyle{ieee_fullname}
\bibliography{egbib}
}
\newpage
$ \ \ \ $
\newpage
\section{Appendix}
Due to space limitations, we omitted including some additional experimental results pertaining to stress tests, which we provide here in Section \ref{sec:stress_test_cont}. We include an additional ablation study on the effect of the hyper-parameter $\tau$, on our robustness metric AUMM, $\delta_{\text{Id}}$, and accuracy in Section \ref{sec:ablation_tau}. In Section~\ref{sec:Aba_RMM}, we report the Rawlsian min-max (RMM) results on identity-related gender and age groups for different methods. %In Section~\ref{sec:trade-offs}, we show the trade-offs between hard and easy samples in terms of identity.%In Section ???? , we report the correlations between CID robust weights with inverse frequency for each identity to evaluate the reliability of the density measure.
Finally, we describe the detailed implementation of ARL, and provide connections and comparisons between CID and baselines in Section~\ref{sec:cc_baselines}.
\subsection{More Results on the stress-tests}
\label{sec:stress_test_cont}
Here we provide additional results to complete the ones provided in Section \ref{sec:stress_test}. In Table \ref{tab:smiling_FNMP}
we provide results on the biased dataset under the FNMP (excluding 90\% of female-negative and male-positive smiling examples) setup. Trends are similar to the experiments included in the main manuscript under Section \ref{sec:stress_test}. We also provide results on the other task of CelebA, namely, Mouth Slightly Open (MSO) in Figure \ref{fig:MSO_stress_test}. As mentioned in the main manuscript, all trends are similar to the smiling task, in terms of CID being consistently more accurate and less sensitive to bias, compared to CE. Also, the gap between the two models increases as bias in the dataset increases. In Table~\ref{tab:smiling_FP_MN_FPMN}, we provide results on the biased CelebA dataset under FP 90\%, MN 90\%, and FPMN 90\% setup. Trends are also consistently similar to the smiling tasks under section~\ref{sec:stress_test}.
In addition, we also add the trends in accuracy and ID-accuracy in the smiling task in Figure \ref{fig:acc_id_acc_smiling_stress_test}.
\begin{table}[htbp]
\centering
\caption{Stress testing the models for the smiling task on the CelebA dataset, by eliminating 90\% of the FNMP subpopulation (90\% of the female-negative samples, and 90\% of the male-positive samples) in the training/validation set. CID achieves the best results on the non-modified test set. This is consistent with other settings reported in Section \ref{sec:stress_test}.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
{\bf FNMP } & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD\\ \hline
CE & 87.61 & 87.28 & 60.47 & 0.1252 & 0.2137 & \underline{0.1479 }\\
IFW & 87.75 & 86.28 & 60.13 & 0.1256 & 0.2161 & 0.1533 \\
DRO & 87.85 & \underline{87.45} &
\underline{60.64} & \underline{0.1240} & \underline{0.2122} & 0.1535 \\
IRM & 87.77 & 86.23 & 60.29 & 0.1256 & 0.2145 & 0.1493 \\
ARL & \underline{88.10} & 86.82 & 59.33 & 0.1285 & 0.2194 & 0.1539 \\\hline
CID & \textbf{88.44} & \textbf{87.93} & \textbf{61.54} & \textbf{0.1224} & \textbf{0.2068} &\textbf{0.1465}\\
\bottomrule
\end{tabular}
}
\label{tab:smiling_FNMP}
\end{table}
\begin{table}[htbp]
\caption{Stree testing on the models for the mouth slightly open (MSO) task on CelebA dataset, by eliminating 90\% FP, MN, FPMN of a subpopulation in the training/validation set. CID achieves the best results.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
\textbf{FP} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 91.70 & 91.29 & 67.58 & 0.1061 & 0.1726 & 0.1349 \\
IFW & \underline{92.44} & \underline{92.09} & \underline{69.53} & \underline{0.1023}
& \underline{0.1612} & \underline{0.1300} \\
DRO & 91.81 & 91.33 & 67.33 & 0.1059 & 0.1707 &0.1323\\
IRM & 91.71 & 91.25 & 67.78 & 0.1083 & 0.1729 & 0.1346 \\
ARL & 91.94 & 91.38 & 67.13 & 0.1071 & 0.1735 & 0.1311 \\\hline
CID & \textbf{92.67} & \textbf{92.16} & \textbf{69.91} & \textbf{0.1000} & \textbf{0.1589} & \textbf{0.1126}\\
\bottomrule
\end{tabular}
}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
\textbf{MN} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 92.36 & 91.67 & 68.33 & 0.1011 & 0.1662 & 0.1200 \\
IFW & \underline{92.67} & \underline{92.22} & \underline{71.17} & \underline{0.0971} & \underline{0.1563} & 0.1265 \\
DRO & 92.54 & 91.89 & 68.37 & 0.1054 & 0.1672 & \underline{0.1166} \\
IRM & 92.62 & 91.96 & 68.46 & 0.1048 & 0.1657 & 0.1192 \\
ARL & 92.34 & 91.86 & 68.59 & 0.1053 & 0.1679 & 0.1213 \\ \hline
CID & \textbf{92.90} & \textbf{92.32} & \textbf{71.87} & \textbf{0.0949} & \textbf{0.1560} & \textbf{0.1161} \\
\bottomrule
\end{tabular}
}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
\textbf{FPMN} & Acc & Id Acc & 10\% Id Acc & $\delta_{\text{Id}}$ & AUMM & AUD \\ \hline
CE & 89.99 & 89.39 & 64.22 & 0.1238 & 0.2011 & 0.1410 \\
IFW & 89.88 & 89.05 & 65.82 & 0.1275 & 0.2033 & 0.1424 \\
DRO & \underline{90.02} & \underline{89.42} & \underline{65.93} & 0.1211 & 0.1988 & \underline{0.1336 } \\
IRM & 89.99 & 89.35 & 65.41 & 0.1212 & \underline{0.1984} & 0.1429 \\
ARL & 90.00 & 89.40 & 63.63 & \underline{ 0.1175} & 0.1987 & 0.1381 \\\hline
CID & \textbf{90.47} & \textbf{ 89.94} & \textbf{67.05} & \textbf{0.1113} & \textbf{0.1883} & \textbf{0.1295} \\ \bottomrule
\end{tabular}
}
\label{tab:smiling_FP_MN_FPMN}
\end{table}
\begin{figure*}[htbp]
\centering
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFP_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMP_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFPMN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFP_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFN_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMP_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMN_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFPMN_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFP_AUMM.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFN_AUMM.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMP_AUMM.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMN_AUMM.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFPMN_AUMM.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFP_std.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFN_std.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMP_std.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMN_std.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFPMN_std.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFP_bottom_10_id_acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFN_bottom_10_id_acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMP_bottom_10_id_acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DMN_bottom_10_id_acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/MSO_DFPMN_bottom_10_id_acc.png}
\caption{Results of different stress tests on the CelebA dataset for the mouth slightly open (MSO) task. We observe trends similar to the ones in the smiling task reported in Figure \ref{fig:OOD_results}. CID seems consistently more robust, and less sensitive to bias compared to CE.}
\label{fig:MSO_stress_test}
\end{figure*}
\subsection{Ablation Studies on $\tau$}
\label{sec:ablation_tau}
In this section, we provide ablation studies of $\tau$ on the smiling task on CelebA, to show its effect on the results of the proposed CID method. As mentioned in the main manuscript in Section \ref{sec:rpse}, and shown in Figure \ref{fig:soft_assignment_weights}, this hyper-parameter implicitly captures the neighborhood radius across which performance is encouraged to remain uniform.
In Figure \ref{fig:tau_ablation}, we report the AUMM, $\delta_{\text{Id}}$, and Accuracy by varying $\tau \in [0.01, 0.9]$ on the default dataset and stressed FP 90\%, MN 90\%, FPMN 90\% datasets. As it can be observed we can see the optimal value (lowest AUMM) is always achieved at 0.1. We hypothesize that as this maps to a specific radius in the embedding space, it captures the optimal neighborhood size for an individual person in the proxy embedding space. This is fairly consistent with the results reported in \cite{ardeshir2022estimating} regarding a specific radius being best for estimating individual fairness.
\begin{figure*}[t!]
\centering
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFP_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DMP_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DMN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFPMN_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFP_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFN_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DMP_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DMN_Id_Acc.png}
\includegraphics[width = 0.19\linewidth]{./figures/Smile_DFPMN_Id_Acc.png}
\caption{Accuracy (Acc) and average per-Id Accuracy (Id Acc) results of different stress tests on the CelebA dataset for the smiling task label. It can be observed that the trend is similar to other metrics in terms of CID being more robust to bias and maintaining higher overall accuracy.}
\label{fig:acc_id_acc_smiling_stress_test}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_default_aumm.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FP90_aumm.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_MN90_aumm.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FPMN90_aumm.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_default_std.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FP90_std.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_MN90_std.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FPMN90_std.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_default_Acc.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FP90_Acc.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_MN90_Acc.png}
\includegraphics[width = 0.24\linewidth]{./figures/tau_smile_FPMN90_Acc.png}
% \includegraphics[width = 0.45\linewidth]{./figures/IMMC_SMILE.jpg}
% \includegraphics[width = 0.45\linewidth]{./figures/IMMC_MSO.jpg}
\vspace{-0.1in}
\caption{The robustness of the proposed CID methods in terms $\tau$. As it can be observed maximum identity-robustness (minimum AUMM and $\delta_{\text{Id}}$), and maximum accuracy are achieved almost consistently at $\tau$=0.1}
\label{fig:tau_ablation}
\end{figure*}
\subsection{Rawlsian Min-Max (RMM) on gender and age}
\label{sec:Aba_RMM}
To complete our experiments, we report fairness in terms of identity-related gender and age attributes. Given that identity groups are defined by a unique set of characteristics, enhancing the fairness of identity can result in improved fairness across gender and age groups.
As gender ($Male$ or not) and age ($Young$ or not) are labeled as binary in CelebA, we evaluate Rawlsian max-min (RMM) to measure disparity.
We compare CID with other baselines and report the results in Table~\ref{tab:gender_age_rmm} under different setups on the Smiling task of CelebA, the improved (reduced) RMM verifies the effectiveness of the CID method on binary identity-related attributes.
% We report the RMM binary-sensitive attribute evaluation metric in terms of identity-related binary attributes gender ($Male$ or not), and age ($Young$ or not) in CelebA.
% Please add the following required packages to your document preamble:
% \usepackage{graphicx}
\begin{table}[h!]
\centering
\caption{Gender and age Rawlsian min-max results on smiling tasks}
\label{tab:gender_age_rmm}
\resizebox{0.4\textwidth}{!}{%
\begin{tabular}{c|c|c
|c|c}\toprule
Gender & Default & MP 90\% & MN 90\% & FPMN 90\% \\ \toprule
CE & 0.01717 & 0.05047 & 0.07859 & 0.05940 \\
IFW & 0.01748 & \underline{ 0.04116 } & \underline{0.05415} & 0.07738 \\
DRO & 0.01671 & 0.04901 & 0.07635 & 0.05933 \\
IRM & \underline{0.01645} & 0.05117 & 0.07792 & \underline{0.05722} \\
ARL & 0.01678 & 0.05366 & 0.07785 & 0.05910 \\\hline
CID & \textbf{ 0.01633} & \textbf{0.03553} & \textbf{0.04893} & \textbf{0.04428} \\
\bottomrule
\end{tabular}%
}
\resizebox{0.4\textwidth}{!}{%
\begin{tabular}{c|c|c
|c|c}\toprule
Age & Default & MP 90\% & MN 90\% & FPMN 90\% \\ \toprule
CE & 0.02251 & 0.04167 & 0.06421 & 0.07548 \\
IFW & 0.02289 & \underline{0.03378} & \underline{0.05536} & 0.08442 \\
DRO & 0.02336 & 0.03838 & 0.06258 & \underline{0.07518} \\
IRM & \underline{0.02251} & 0.04001 & 0.06556 & 0.07698 \\
ARL & 0.02271 & 0.04136 & 0.06744 &0.07624\\ \hline
CID & \textbf{0.02231} & \textbf{0.03131} & \textbf{0.05288} & \textbf{0.06411} \\
\bottomrule
\end{tabular}%
}
\end{table}
% \begin{table}[h]
% \centering
% \caption{Correlations between cid robust weights with the inverse frequency of positive and negative samples for each identity sample to improve identity robustness}
% \resizebox{0.48\textwidth}{!}{
% \begin{tabular}{c|c|c|c|c} \toprule
% & Default &MP 90\% & MN 90\%& FPMN 90\% \\ \hline
% Pearson Cor &0.7276 &0.9511 & 0.9252 &0.8081 \\
% p-value &5.213e-06 &8.1549e-16 & 2.668e-13&6.6571e-08 \\
% \bottomrule
% \end{tabular}
% }
% \label{tab:cor_cid_inverse_frequency}
% \end{table}
% \subsection{Trade-offs between hard and easy identities}
% % \label{sec:trade-offs}
% There might be a trade-off in some instances. We added more buckets to Table \ref{tab:id_trade_offs} to confirm this. There is less of a trade-off in more biased scenarios (no trade-off for FPMN90\%). But the improvement margins of CID over CE on different buckets are decreased from tail to head identities.
% \begin{table}[h]
% \centering
% \caption{ The trade-off between easy and hard samples.}
% \resizebox{0.45\textwidth}{!}{%
% \begin{tabular}{c|c|c|c|c|c}\toprule
% & Id Acc & Bottom 25\% & Bottom 50\% & Top 50\% & Top 25\% \\ \toprule
% \multirow{2}{*}{Default} & CE & 77.42 & 85.45 & \textbf{98.75} & 1 \\
% & CID &\textbf{77.66} & \textbf{85.61} & 98.59 & 1 \\ \hline
% \multirow{2}{*}{FPMN 90\%} & CE & 65.87 & 77.05 & 95.84 & 1 \\
% & CID &\textbf{68.86} & \textbf{78.00} &\textbf{ 96.14 } & 1
% \\ \bottomrule
% \end{tabular}}%
% \label{tab:id_trade_offs}
% \vspace{-0.1in}
% \end{table}
\subsection{Connects and Comparison with other baselines}
\label{sec:cc_baselines}
In this section, we draw parallels and provide comparisons of CID with the methods that can be broadly categorized into the sample weighting scheme including ARL, DRO, and IFW class-balancing method.
\subsubsection{Adversarially Reweighted Learning (ARL)}
ARL~\cite{lahoti2020fairness} is a fairness-aware method that designed to handle fairness under awareness challenges. To achieve this, ARL makes use of adversarial training to emphasize more on the samples
that make significant errors and improve improve accuracy for the protected groups. Specifically, ARL optimizes the following min-max objective:
\begin{align}
\label{eqn:appendix_ARL}
\min_\w\max_{\phi} \lambda_{\phi}(\x_i, y_i) \ell_i(\w;\x_i,y_i)
\end{align}
where $\w$ is the model (learner) parameters, and $\phi$ denotes the adversarial network, $f_\phi$ represents the output of adversarial network
and $\lambda_{\phi}(\x_i,y_i) = 1/n + \frac{f_{\phi}(\x_i,y_i)}{\sum_{i=1}^n f_{\phi}(\x_i,y_i)}$ is the adversarial assignment of the weight vector $\lambda_\phi: f_{\phi} \rightarrow R$ so as to maximize the expected loss and $f_\phi: X\times Y\rightarrow R$ is the output of the adversary network designed to identify the regions where the model makes significant errors.
In our case, we aim to balance the positive and negative samples for each identity during the training and emphasize more on the minorities to achieve fairness under unawareness.
As the same facial features close to each other in the face recognition space, we make use of face recognition embeddings as proxy vectors of identities due to the unavailable of ground truth label information. We assign conditional inverse density robust weight $p^\tau_i$ to each sample defined in the neighborhood of the proxy space to balance the positive and negative samples that share the same facial features, i.e, identity.
\begin{align}
\min_\w \sum\limits_{i=1}^n\frac{p^\tau_i}{Z_{y_i}} \ell_i(\w;\x_i,y_i)
\end{align}
where $n$ represents the size of $\D$, i.e, $n = |\D|$, $ p_{i}^\tau = \exp(\frac{\z_i^\top\z_i}{\tau})/ \sum_{k=1}^{|\D_{y_i}|}\exp(\frac{\z_i^\top\z_k}{\tau})$ and $\tau$ represents the size of the neighborhood within which samples are considered to be from the same identity group. $Z_{y_i} = \sum_{j\in \D_{y_i}}p_j^{\tau}$ to balance class contributions.
Hence by replace $\lambda_\phi(\x_i,y_i)$ with $p_i^\tau$ in Euqation~(\ref{eqn:appendix_ARL}), ARL share the same objective with our proposed method.
Compare with $p_i^\tau$, $\lambda_\phi(\x_i,y_i)$ is obtained via maximizing the objectives combined with the learner prediction losses, i.e, $\max_{\phi} \lambda_{\phi}(\x_i, y_i) \ell_i(\w;\x_i,y_i)$.
Hence, the $\lambda_{\phi}(\x_i, y_i)$ is obtained based on the sample loss predictions rather than features that represent group information such as identity, which leads to poorer performance of ARL than CID in terms of identity fairness, as observed in all experiments
\noindent\textbf{ARL Implementation}
The same as other baselines, the learner is implemented using ResNet18. For the adversarial head, we implement a linear model with sigmoid activation function where the input is the representation feature learned from learner. The learning rate of adversarial learner $\phi$ is tuned in $\{1e$-$2, 1e$-$3, 1e$-$4\}$.
\subsubsection{Distributionally Robust Learning}
DRO~\cite{li2021tilted, qi2020attentional} has been show an effective methods to improve the worst group accuracy such that the fairness can be improved.
Similar to ARL, DRO
\begin{align}
\min\limits_{\w}\max\limits_{\p\in\U} p_i\ell_i(\w;\x_i,y_i)
\end{align}
in which $\p$ is defined in the uncertainty set $\U$.
Hence, after getting $\p^*$ by solving the inner maximization problem, the DRO reduces to a instance sample weighting method that
\begin{align}
\min\limits_{\w} p^*_i\ell_i(\w;\x_i,y_i)
\end{align}
To be more specific, when $\U = \{\p\in\Delta, \sum\limits_{i=1}^n p_i\log n p_i \leq \rho\}$ in ~\cite{li2021tilted, qi2020attentional}, where $\p\in\Delta:\sum\limits_{i=1}^n p_i = 1$, $\sum\limits_{i=1}^n p_i\log n p_i \leq\rho$ represents the KL-divergence discrepancy constraint between $\p$ and uniform sampling weights $\frac{\textbf{1}}{n}, 1\in\R^n$. Given this uncertainty set, $p_i^*$ has the close form of
$p_i^* = \exp(\frac{\ell_i(\w;\x_i,y_i)}{\nu})/\sum\limits_{i=1}^{n}\exp(\frac{\ell_i(\w;\x_i,y_i)}{\nu})$, where $\nu$ is a Lagrange multiplier introduced by $\U$.
Therefore, we can see the instance weights $p_i^*$ for each sample introduced by DRO is proportional to the loss scales, while our fairness robust weights $p_i^{\tau}$ defined on the pairwise similarity in the neighborhood face proxy embeddings of sample $(\x_i,y_i)$ to balance the contribution of positive and negative classes for any arbitrary facial features that are considered belonging to the same identity.
\subsubsection{Inverse Frequency Weighting (IFW)}
IFW~\cite{huang2016learning, wang2017learning} aims to balance the contributions of different classes by assigning class-balancing weights to each sample that is proportional to the number of samples for each class, i.e,
\begin{align}
\sum\limits_{i=1}^n p_i\ell_{i}(\w;\x_i,y_i)
\end{align}
where $p_i \propto \frac{1}{N_{y_i}}$, where $N_{y_i}$ represents the frequency of class $y_i$.
As we explained in Section~\ref{sec:rpse}, our assign robust weights $p_i^\tau\rightarrow \frac{1}{|\D_{y_i}|}$ when $\tau \rightarrow \infty$ such that CID reduces to IFW.
Hence, IFW is a special case of CID. When $\tau \in (0,\infty)$, CID considers more fine-grained rarity of local neighborhood face proxy embedding, i.e, $p_i^\tau $ balances the positive and negative samples within each local neighborhood which serves as a proxy of identity, to improve identity robustness.
% \begin{table*}[]
% \caption{OOD Smiling attribute biased group distribution results in part 2.}
% \label{tab:my-table}
% \resizebox{.95\textwidth}{!}{
% \begin{tabular}{c||cc|cc|cc||cc|cc|cc}
% % random seed: 1, 3, 5, 7
% % batch size 256
% \toprule
% \multicolumn{1}{c|}{Delete Percentage} & \multicolumn{6}{c|}{90\%} & \multicolumn{6}{c}{75\%} \\ \hline
% \multicolumn{1}{c|}{Groups} & \multicolumn{2}{c|}{DFS} & \multicolumn{2}{c|}{DMN} & \multicolumn{2}{c|}{DFSMN} & \multicolumn{2}{c|}{DFS} & \multicolumn{2}{c|}{DMN} & \multicolumn{2}{c}{DFSMN} \\ \hline
% Methods & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID \\ \midrule
% Test Acc & 0.9038 & \textbf{0.9131} & 0.9015 & \textbf{ 0.9115} & 0.8519 & \textbf{0.9144} & 0.9121 & {\bf 0.9179} & 0.9126 & {\bf 0.9181} & 0.8927 & {\bf 0.9176 } \\
% ID Acc & 0.8972 & \textbf{0.9073} & 0.8942 & \textbf{0.9036} & 0.8507 & \textbf{0.9076 } & 0.9035 & \textbf{0.9113} & 0.9061 & {\bf 0.9102 } & 0.8831 & {\bf 0.9103} \\
% 10\% ID Acc & 0.6596 & \textbf{0.6829} & 0.6322 & \textbf{ 0.6615} & 0.5585 & \textbf{0.6745} & 0.6670 & {\bf 0.6926} & 0.6639 & {\bf 0.6804 } & 0.6382 & {\bf0.6839 } \\
% Worst Group Acc & 0.8847 & \textbf{0.9228} & 0.7739 & \textbf{ 0.8294 } & 0.7166 & \textbf{ 0.8970 } & 0.8753 & {\bf 0.9132} & 0.8409 & {\bf 0.8837} & 0.9255 & {\bf 0.9191 } \\
% Lst Group Acc & 0.8296 & \textbf{0.8793} & 0.7739 & \textbf{ 0.8294 } & 0.7158 & \textbf{ 0.8967 } & 0.8665 & {\bf 0.8996} & 0.8409 & {\bf 0.8837} & 0.7584 & {\bf 0.8813} \\
% F-Acc & 0.8971 & \textbf{0.9180} & \underline{0.9307} & \textbf{ 0.9308 } & 0.8694 & \textbf{ 0.9227 } & 0.9129 & {\bf 0.9248} & {\bf 0.9303} & 0.9295 & 0.9048 & {\bf 0.9294 } \\
% M-Acc & \textbf{0.9147} & 0.9054 & 0.8544 & \textbf{0.8803} & 0.8237 & \textbf{ 0.9012 } & {\bf 0.9110} & 0.9068 & 0.8842 &{\bf 0.8999 } & 0.8731 & {\bf 0.9058 } \\ \midrule
% ID Acc Std & 0.1106 & \textbf{0.1033} & 0.1175 & \textbf{ 0.1083 } & 0.1357 & \textbf{0.1053} & 0.1086 & {\bf 0.1011} & 0.1082 & {\bf 0.1035} & 0.1147 & {\bf 0.1033} \\
% AUMM & 0.1848 & \textbf{0.1701} & 0.1930 & \textbf{0.1786} & 0.2373 & \textbf{ 0.1716 } & 0.1777 & {\bf 0.1651} & 0.1758 & {\bf 0.1698} & 0.1943 & {\bf 0.1682} \\
% \bottomrule
% % Gender RMM & 0.01572 & 0.01906 & 0.02709 & 0.02132 & 0.02138 & 0.02312 & 0.00284 & 0.01944 & 0.04956 & {\bf 0.03184} & 0.03463 & {\bf 0.02072 } \\ \hline
% %Best Epoch & 20 & 20 & 20 & 20 & 20 & 20 & 19 & 19 & 18 & 18 & 17 & 17 \\ \hline
% \end{tabular}}
% \end{table*}
% \begin{table*}[]
% \centering
% \caption{OOD Smiling biased DFN and DMS group distributioin results }
% \label{tab:OOD_DFN_DMS}
% \resizebox{.95\textwidth}{!}{
% \begin{tabular}{c||cc|cc||cc|cc||cc|cc||cc|cc}
% \toprule
% \multicolumn{1}{c|}{Delete Percentage} & \multicolumn{4}{c|}{25\%} & \multicolumn{4}{c|}{50\%} & \multicolumn{4}{c|}{75\%} & \multicolumn{4}{c}{90\%} \\ \hline
% \multicolumn{1}{c|}{Groups} & \multicolumn{2}{c|}{FN} & \multicolumn{2}{c|}{MP} & \multicolumn{2}{c|}{FN} & \multicolumn{2}{c|}{MP} & \multicolumn{2}{c|}{FN} & \multicolumn{2}{c|}{MP} & \multicolumn{2}{c|}{FN} & \multicolumn{2}{c}{MP} \\ \hline
% Methods & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID \\ \midrule
% Test Acc & 0.9250 & 0.9262 & 0.9258 & 0.9261 & 0.9224 & 0.9245 & 0.9252 & 0.9251 & 0.9117 & 0.9193 & 0.9196 & 0.9228 & 0.9038 & 0.9131 & 0.9121 & 0.9161 \\
% ID Acc & 0.9186 & 0.9201 & 0.9180 & 0.919 & 0.9163 & 0.9186 & 0.9175 & 0.9187 & 0.9079 & 0.9134 & 0.9109 & 0.9143 & 0.8972 & 0.9073 & 0.9010 & 0.9060 \\
% 10\% ID Acc & 0.7125 & \textbf{0.7129} & 0.7074 & \textbf{0.7063 } & 0.7016 & \textbf{0.7097} & 0.7023 & \textbf{0.7072} & 0.6884 & \textbf{0.7040 } & 0.6814 & \textbf{0.6899} & 0.6596 &\textbf{ 0.6829} & 0.6410 & \textbf{0.6605} \\
% Worst Group Acc & 0.9093 & 0.9175 & 0.8755 & 0.8871 & 0.8870 & 0.9043 & 0.8481 & 0.8682 & 0.8375 & 0.8717 & 0.7936 & 0.8276 & 0.8848 & 0.9228 & 0.7343 & 0.7794 \\
% Lst Group Acc & 0.8944 & 0.8938 & 0.8755 & 0.8871 & 0.8818 & 0.8882 & 0.8481 & 0.8682 & 0.8375 & 0.8654 & 0.7936 & 0.8276 & 0.8296 & 0.8793 & 0.7343 & 0.7794 \\
% F-Acc & 0.9306 & 0.9324 & 0.9315 & 0.9319 & 0.9262 & 0.9287 & 0.9313 & 0.9316 & 0.9100 & 0.9219 & 0.9296 & 0.9309 & 0.8971 & 0.9180 & 0.9303 & 0.9282 \\
% M-Acc & 0.9155 & 0.9166 & 0.9163 & 0.9175 & 0.9164 & 0.9179 & 0.9145 & 0.9149 & 0.9146 & 0.9153 & 0.9036 & 0.9098 & 0.9147 & 0.9054 & 0.8830 & 0.8963 \\ \midrule
% ID Acc Std & 0.0957 & \textbf{0.0944} & 0.0993 & \textbf{0.0973} & 0.0996 & \textbf{0.0967 } & 0.0992 & \textbf{0.0973} & 0.1022 & \textbf{0.0979} & 0.1047 & \textbf{0.1024} & 0.1106 & \textbf{0.1033} & 0.1168 & \textbf{0.1092} \\
% AUMM & 0.1542 &\textbf{ 0.1533 } & 0.1569 & \textbf{0.1567} & 0.1608 & \textbf{0.1561} & 0.1600 & \textbf{0.1568} & 0.1674 & \textbf{0.1600} & 0.1698 & \textbf{0.1646} & 0.1848 & \textbf{0.1701} & 0.1912 & \textbf{0.1790} \\
% \bottomrule
% \end{tabular}}
% \end{table*}
% \begin{table*}[]
% \caption{MSO OOD 1}
% \label{tab:my-table}
% \resizebox{.95\textwidth}{!}{\begin{tabular}{c|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc}
% \toprule
% Delete Percentage & \multicolumn{10}{c|}{75\%} & \multicolumn{10}{c}{90\%} \\ \midrule
% & MP & & \multicolumn{2}{c}{MN} & \multicolumn{2}{c}{FP} & \multicolumn{2}{c}{FN} & \multicolumn{2}{c|}{FPMN} & \multicolumn{2}{c}{MP} & \multicolumn{2}{c}{MN} & \multicolumn{2}{c}{FP} & \multicolumn{2}{c}{FN} & \multicolumn{2}{c}{FPMN} \\ \midrule
% & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID \\ \midrule
% Test Acc & 0.9334 & \textbf{0.9376} & 0.9339 & \textbf{0.9354} & 0.9202 & \textbf{0.9342} & 0.9351 & \textbf{0.9372} & 0.9254 & \textbf{0.9380} & 0.9257 & \textbf{0.9322} & 0.9236 & \textbf{0.9290} & 0.9170 & 0.9267 & 0.9208 & \textbf{0.9318} & 0.8867 & \textbf{0.9329} \\
% ID Acc & 0.9276 & \textbf{0.9333} & 0.9279 & \textbf{0.9293} & 0.9231 & \textbf{0.9293} & 0.9322 & \textbf{0.9333} & 0.9202 & \textbf{0.9325} & 0.9197 & \textbf{0.9256} & 0.9167 & \textbf{0.9231} & 0.9129 & 0.9216 & 0.9188 & \textbf{0.9286} & 0.8821 & \textbf{0.9292} \\
% 10\% ID Acc & 0.7121 & \textbf{0.7290} & 0.7215 & \textbf{0.7219} & 0.7013 & \textbf{0.7214} & 0.7299 & \textbf{0.7339} & 0.7075 & \textbf{0.7311} & 0.6829 & \textbf{0.7024} & 0.6833 & \textbf{0.7058} & 0.6730 & 0.6991 & 0.6953 & \textbf{0.7212} & 0.6063 & \textbf{0.7268} \\
% Worst Group Acc & 0.8580 & \textbf{0.8869} & 0.9002 & \textbf{0.9242} & 0.9199 & \textbf{0.9427} & 0.8878 & \textbf{0.9079} & 0.8880 & \textbf{0.9479} & 0.8080 & \textbf{0.8457} & 0.8464 & \textbf{0.8903} & 0.9166 & 0.9415 & 0.8266 & \textbf{0.8760} & 0.7904 & \textbf{0.9339} \\
% Lst Group Acc & 0.8580 & \textbf{0.8869} & 0.9002 & \textbf{0.9242} & 0.9199 & \textbf{0.9427} & 0.8878 & \textbf{0.9079} & 0.8773 & \textbf{0.9271} & 0.8080 & \textbf{0.8457} & 0.8464 & \textbf{0.8903} & 0.8329 & 0.8772 & 0.8266 & \textbf{0.8760} & 0.7861 & \textbf{0.9260} \\
% F-Acc & 0.9373 & \textbf{0.9391} & 0.9381 & \textbf{0.9356} & 0.9325 & \textbf{0.9313} & 0.9320 & \textbf{0.9348} & 0.9241 & \textbf{0.9390} & 0.9356 & \textbf{0.9383} & 0.8464 & \textbf{0.8903} & & & 0.9095 & \textbf{0.9271} & 0.8956 & \textbf{0.9340} \\
% M-Acc & 0.9273 & \textbf{0.9347} & 0.9275 & \textbf{0.9351} & 0.9384 & \textbf{0.9339} & 0.9402 & \textbf{0.9411} & 0.9280 & \textbf{0.9375} & 0.9099 & \textbf{0.9226} & 0.9014 & \textbf{0.9213} & & & 0.9393 & \textbf{0.9395} & 0.8738 & \textbf{0.9313} \\ \midrule
% ID Acc Std & 0.0958 & \textbf{0.0900} & 0.0947 & \textbf{0.0942} & 0.0996 & 0.0934 & 0.0889 & \textbf{0.0877} & 0.0952 & \textbf{0.0886} & 0.0985 & \textbf{0.1033} & 0.1027 & \textbf{0.0973} & 0.1069 & 0.1000 & 0.0986 & \textbf{0.0914} & 0.2099 & \textbf{0.1449} \\
% AUMM & 0.1513 & \textbf{0.1427} & 0.1456 & \textbf{0.1490} & 0.1593 & \textbf{0.1470} & 0.1409 & \textbf{0.1390} & 0.1563 & \textbf{0.1409} & 0.1652 & \textbf{0.1542} & 0.1662 & \textbf{0.1560} & 0.1726 & 0.1589 & 0.1616 & \textbf{0.1461} & 0.1238 & \textbf{0.0904} \\
% \bottomrule
% \end{tabular}}
% \end{table*}
% Figure~\ref{fig:identity_distribution} shows the identity distribution on these manipulated versions of the dataset.
% \begin{figure}[htbp]
% \centering
% \includegraphics[width = 0.48\linewidth]{./figures/id_val_ID_OOD.png}
% \includegraphics[width = 0.48\linewidth]{./figures/id_test_ID.png}
% \caption{Identity ID and OOD distribution on validation and test data.}
% \label{fig:identity_distribution}
% \end{figure}
% \begin{table*}[]
% \caption{MSO OOD 2}
% \label{tab:my-table}
% \centering
% \resizebox{.95\textwidth}{!}{
% \begin{tabular}{c|cc|cc|cc|cc|cc|cc|cc|cc|cc|cc}
% \toprule
% Delete Percentage & \multicolumn{10}{c|}{25} & \multicolumn{10}{c}{50} \\ \hline
% & \multicolumn{2}{c}{MP} & \multicolumn{2}{c}{MN} & \multicolumn{2}{c}{FP} & \multicolumn{2}{c}{FN} & \multicolumn{2}{c|}{FPMN} & \multicolumn{2}{c}{MP} & \multicolumn{2}{c}{MN} & \multicolumn{2}{c}{FP} & \multicolumn{2}{c}{FN} & \multicolumn{2}{c}{FPMN} \\ \hline
% & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID & CE & CID \\ \hline
% Test Acc & 0.9399 & \textbf{0.9406} & 0.9405 & \textbf{0.9398} & 0.9396 & 0.9406 & 0.9410 & \textbf{0.9415} & 0.9386 & 0.9392 & Y & \textbf{N} & 0.9387 & \textbf{0.9393} & 0.9355 & 0.9376 & N & \textbf{Y} & 0.9357 & \textbf{0.9391} \\
% ID Acc & 0.9399 & \textbf{0.9406} & 0.9351 & \textbf{0.9361} & 0.9398 & \textbf{0.9401} & 0.9365 & \textbf{0.9372} & 0.9331 & 0.9340 & & \textbf{} & 0.9338 & \textbf{0.9349} & 0.9298 & 0.9393 & & \textbf{} & 0.9322 & \textbf{0.9345} \\
% 10\% ID Acc & 0.9361 & \textbf{0.7386} & 0.7491 & \textbf{0.7395} & 0.7369 & \textbf{0.7382} & 0.7403 & \textbf{0.7502} & 0.7263 & \textbf{0.7312} & & \textbf{} & 0.7352 & \textbf{0.7411} & 0.7143 & 0.7295 & & \textbf{} & 0.7366 & \textbf{0.7401} \\
% \hline
% Worst Group Acc & 0.9267 & \textbf{0.9264} & 0.9476 & \textbf{0.9491} & 0.9245 & \textbf{0.9324} & 0.9301 & \textbf{0.9357} & & \textbf{} & & \textbf{} & 0.9296 & \textbf{0.9429} & 0.9401 & 0.9459 & & \textbf{} & 0.9122 & \textbf{0.9394} \\
% Lst Group Acc & 0.9207 & \textbf{0.9264} & 0.9367 & \textbf{0.9333} & 0.9289 & 0.9370 & 0.9278 & \textbf{0.9300} & 0.9213 & \textbf{0.9372} & & \textbf{} & & \textbf{0.9316} & 0.9070 & 0.9233 & & \textbf{} & 0.9271 & \textbf{0.9480} \\
% F-Acc & 0.9386 & \textbf{0.9394} & 0.939 & \textbf{0.9391} & & \textbf{} & 0.9399 & \textbf{0.9499} & & \textbf{} & & \textbf{} & \textbf{0.9392} & 0.9377 & 0.9421 & 0.9354 & & \textbf{} & 0.9364 & \textbf{0.9386} \\
% M-Acc & 0.9423 & \textbf{0.9429} & 0.9428 & \textbf{0.9439} & & \textbf{} & 0.9426 & \textbf{0.9427} & & \textbf{} & & \textbf{} & 0.9380 & \textbf{0.9421} & 0.9451 & 0.9412 & & \textbf{} & 0.9361 & \textbf{0.9404} \\ \hline
% ID Acc Std & 0.0887 & \textbf{0.0870} & 0.0881 & \textbf{0.0865} & 0.0875 & 0.0866 & 0.0863 & \textbf{0.0827} & 0.09214 & \textbf{0.08931} & & \textbf{} & 0.0878 & \textbf{0.0873} & 0.0957 & 0.09135 & & \textbf{} & 0.0867 & \textbf{0.0864} \\
% AUMM & 0.1383 & \textbf{0.1365} & 0.1377 & \textbf{0.1364} & 0.1379 & 0.1336 & 0.1348 & \textbf{0.1313} & 0.1430 & 0.1409 & & \textbf{} & 0.1387 & \textbf{0.1380} & 0.1486 & 0.1421 & & \textbf{} & 0.1379 & \textbf{0.137} \\
% \bottomrule
% \end{tabular}}
% \end{table*}
\end{document}
\begin{thebibliography}{10}\itemsep=-1pt
\bibitem{adragna2020fairness}
Robert Adragna, Elliot Creager, David Madras, and Richard Zemel.
\newblock Fairness and robustness in invariant learning: A case study in
toxicity classification.
\newblock {\em arXiv preprint arXiv:2011.06485}, 2020.
\bibitem{albiero2021img2pose}
Vitor Albiero, Xingyu Chen, Xi Yin, Guan Pang, and Tal Hassner.
\newblock img2pose: Face alignment and detection via 6dof, face pose
estimation.
\newblock In {\em Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition}, pages 7617--7627, 2021.
\bibitem{alldieck2018detailed}
Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard
Pons-Moll.
\newblock Detailed human avatars from monocular video.
\newblock In {\em 2018 International Conference on 3D Vision (3DV)}, pages
98--109. IEEE, 2018.
\bibitem{ardeshir2022estimating}
Shervin Ardeshir, Cristina Segalin, and Nathan Kallus.
\newblock Estimating structural disparities for face models.
\newblock In {\em Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition}, pages 10358--10367, 2022.
\bibitem{arjovsky2019invariant}
Martin Arjovsky, L{\'e}on Bottou, Ishaan Gulrajani, and David Lopez-Paz.
\newblock Invariant risk minimization.
\newblock {\em arXiv preprint arXiv:1907.02893}, 2019.
\bibitem{bechavod2017penalizing}
Yahav Bechavod and Katrina Ligett.
\newblock Penalizing unfairness in binary classification.
\newblock {\em arXiv preprint arXiv:1707.00044}, 2017.
\bibitem{ben2013robust}
Aharon Ben-Tal, Dick Den~Hertog, Anja De~Waegenaere, Bertrand Melenberg, and
Gijs Rennen.
\newblock Robust solutions of optimization problems affected by uncertain
probabilities.
\newblock {\em Management Science}, 59(2):341--357, 2013.
\bibitem{beutel2019fairness}
Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt,
Zhe Zhao, Lichan Hong, Ed~H Chi, et~al.
\newblock Fairness in recommendation ranking through pairwise comparisons.
\newblock In {\em Proceedings of the 25th ACM SIGKDD international conference
on knowledge discovery \& data mining}, pages 2212--2220, 2019.
\bibitem{bottou2012stochastic}
L{\'e}on Bottou.
\newblock Stochastic gradient descent tricks.
\newblock In {\em Neural networks: Tricks of the trade}, pages 421--436.
Springer, 2012.
\bibitem{caton2020fairness}
Simon Caton and Christian Haas.
\newblock Fairness in machine learning: A survey.
\newblock {\em arXiv preprint arXiv:2010.04053}, 2020.
\bibitem{cui2019class}
Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie.
\newblock Class-balanced loss based on effective number of samples.
\newblock In {\em Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition}, pages 9268--9277, 2019.
\bibitem{diana2021minimax}
Emily Diana, Wesley Gill, Michael Kearns, Krishnaram Kenthapadi, and Aaron
Roth.
\newblock Minimax group fairness: Algorithms and experiments.
\newblock In {\em Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics,
and Society}, pages 66--76, 2021.
\bibitem{du2021fairness}
Mengnan Du, Subhabrata Mukherjee, Guanchu Wang, Ruixiang Tang, Ahmed Awadallah,
and Xia Hu.
\newblock Fairness via representation neutralization.
\newblock {\em Advances in Neural Information Processing Systems},
34:12091--12103, 2021.
\bibitem{elazar2018adversarial}
Yanai Elazar and Yoav Goldberg.
\newblock Adversarial removal of demographic attributes from text data.
\newblock {\em arXiv preprint arXiv:1808.06640}, 2018.
\bibitem{hashimoto2018fairness}
Tatsunori Hashimoto, Megha Srivastava, Hongseok Namkoong, and Percy Liang.
\newblock Fairness without demographics in repeated loss minimization.
\newblock In {\em International Conference on Machine Learning}, pages
1929--1938. PMLR, 2018.
\bibitem{hebert2018multicalibration}
Ursula H{\'e}bert-Johnson, Michael Kim, Omer Reingold, and Guy Rothblum.
\newblock Multicalibration: Calibration for the (computationally-identifiable)
masses.
\newblock In {\em International Conference on Machine Learning}, pages
1939--1948. PMLR, 2018.
\bibitem{huang2016learning}
Chen Huang, Yining Li, Chen~Change Loy, and Xiaoou Tang.
\newblock Learning deep representation for imbalanced classification.
\newblock In {\em Proceedings of the IEEE conference on computer vision and
pattern recognition}, pages 5375--5384, 2016.
\bibitem{huang2019facial}
Yunxin Huang, Fei Chen, Shaohe Lv, and Xiaodong Wang.
\newblock Facial expression recognition: A survey.
\newblock {\em Symmetry}, 11(10):1189, 2019.
\bibitem{facerecognitionmodel}
Davis King.
\newblock https://github.com/davisking/dlib-models.
\bibitem{kumar2009attribute}
Neeraj Kumar, Alexander~C Berg, Peter~N Belhumeur, and Shree~K Nayar.
\newblock Attribute and simile classifiers for face verification.
\newblock In {\em 2009 IEEE 12th international conference on computer vision},
pages 365--372. IEEE, 2009.
\bibitem{lahoti2020fairness}
Preethi Lahoti, Alex Beutel, Jilin Chen, Kang Lee, Flavien Prost, Nithum Thain,
Xuezhi Wang, and Ed Chi.
\newblock Fairness without demographics through adversarially reweighted
learning.
\newblock {\em Advances in neural information processing systems}, 33:728--740,
2020.
\bibitem{last1711oversampling}
F Last, G Douzas, and F Bacao.
\newblock Oversampling for imbalanced learning based on k-means and smote.
arxiv 2017.
\newblock {\em arXiv preprint arXiv:1711.00837}.
\bibitem{li2021tilted}
Tian Li, Ahmad Beirami, Maziar Sanjabi, and Virginia Smith.
\newblock On tilted losses in machine learning: Theory and applications.
\newblock {\em arXiv preprint arXiv:2109.06141}, 2021.
\bibitem{liu2021just}
Evan~Z Liu, Behzad Haghgoo, Annie~S Chen, Aditi Raghunathan, Pang~Wei Koh,
Shiori Sagawa, Percy Liang, and Chelsea Finn.
\newblock Just train twice: Improving group robustness without training group
information.
\newblock In {\em International Conference on Machine Learning}, pages
6781--6792. PMLR, 2021.
\bibitem{liu2015faceattributes}
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang.
\newblock Deep learning face attributes in the wild.
\newblock In {\em Proceedings of International Conference on Computer Vision
(ICCV)}, December 2015.
\bibitem{park2022fair}
Sungho Park, Jewook Lee, Pilhyeon Lee, Sunhee Hwang, Dohyung Kim, and Hyeran
Byun.
\newblock Fair contrastive learning for facial attribute classification.
\newblock In {\em Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition}, pages 10389--10398, 2022.
\bibitem{qi2020attentional}
Qi Qi, Yi Xu, Rong Jin, Wotao Yin, and Tianbao Yang.
\newblock Attentional biased stochastic gradient for imbalanced classification.
\newblock {\em arXiv preprint arXiv:2012.06951}, 2020.
\bibitem{rawls2001justice}
John Rawls.
\newblock {\em Justice as fairness: A restatement}.
\newblock Harvard University Press, 2001.
\bibitem{rieger2020interpretations}
Laura Rieger, Chandan Singh, William Murdoch, and Bin Yu.
\newblock Interpretations are useful: penalizing explanations to align neural
networks with prior knowledge.
\newblock In {\em International conference on machine learning}, pages
8116--8126. PMLR, 2020.
\bibitem{sagawa2019distributionally}
Shiori Sagawa, Pang~Wei Koh, Tatsunori~B Hashimoto, and Percy Liang.
\newblock Distributionally robust neural networks for group shifts: On the
importance of regularization for worst-case generalization.
\newblock {\em arXiv preprint arXiv:1911.08731}, 2019.
\bibitem{wadsworth2018achieving}
Christina Wadsworth, Francesca Vera, and Chris Piech.
\newblock Achieving fairness through adversarial learning: an application to
recidivism prediction.
\newblock {\em arXiv preprint arXiv:1807.00199}, 2018.
\bibitem{wang2019symmetric}
Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey.
\newblock Symmetric cross entropy for robust learning with noisy labels.
\newblock In {\em Proceedings of the IEEE/CVF International Conference on
Computer Vision}, pages 322--330, 2019.
\bibitem{wang2017learning}
Yu-Xiong Wang, Deva Ramanan, and Martial Hebert.
\newblock Learning to model the tail.
\newblock {\em Advances in neural information processing systems}, 30, 2017.
\bibitem{wu2019facial}
Yue Wu and Qiang Ji.
\newblock Facial landmark detection: A literature survey.
\newblock {\em International Journal of Computer Vision}, 127(2):115--142,
2019.
\bibitem{yan2020fair}
Shen Yan, Hsien-te Kao, and Emilio Ferrara.
\newblock Fair class balancing: Enhancing model fairness without observing
sensitive attributes.
\newblock In {\em Proceedings of the 29th ACM International Conference on
Information \& Knowledge Management}, pages 1715--1724, 2020.
\bibitem{zhang2018mitigating}
Brian~Hu Zhang, Blake Lemoine, and Margaret Mitchell.
\newblock Mitigating unwanted biases with adversarial learning.
\newblock In {\em Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics,
and Society}, pages 335--340, 2018.
\bibitem{zhangfairness}
Fengda Zhang, Kun Kuang, Long Chen, Yuxuan Liu, Chao Wu, and Jun Xiao.
\newblock Fairness-aware contrastive learning with partially annotated
sensitive attributes.
\newblock In {\em The Eleventh International Conference on Learning
Representations}.
\bibitem{zhang2015learning}
Zhanpeng Zhang, Ping Luo, Chen-Change Loy, and Xiaoou Tang.
\newblock Learning social relation traits from face images.
\newblock In {\em Proceedings of the IEEE International Conference on Computer
Vision}, pages 3631--3639, 2015.
\end{thebibliography}