[Paper Review] Fair SA: Sensitivity Analysis for Fairness in Face Recognition

Aparna Joshi, Xavier Suau+3 more|arXiv (Cornell University)|Feb 8, 2022

Face Recognition and Perception4 citations

TL;DR

This paper proposes Fair SA, a novel framework that extends Visual Psychophysics Sensitivity Analysis (VPSA) to evaluate group fairness in face recognition models under image perturbations. By measuring model robustness and subgroup-specific bias across incremental degradations—such as motion blur, exposure, and JPEG compression—it reveals that even high-accuracy models exhibit hidden unfairness under real-world distortions, with VGGFace2 showing the least fairness at high perturbation levels despite strong robustness.

ABSTRACT

As the use of deep learning in high impact domains becomes ubiquitous, it is increasingly important to assess the resilience of models. One such high impact domain is that of face recognition, with real world applications involving images affected by various degradations, such as motion blur or high exposure. Moreover, images captured across different attributes, such as gender and race, can also challenge the robustness of a face recognition algorithm. While traditional summary statistics suggest that the aggregate performance of face recognition models has continued to improve, these metrics do not directly measure the robustness or fairness of the models. Visual Psychophysics Sensitivity Analysis (VPSA) [1] provides a way to pinpoint the individual causes of failure by way of introducing incremental perturbations in the data. However, perturbations may affect subgroups differently. In this paper, we propose a new fairness evaluation based on robustness in the form of a generic framework that extends VPSA. With this framework, we can analyze the ability of a model to perform fairly for different subgroups of a population affected by perturbations, and pinpoint the exact failure modes for a subgroup by measuring targeted robustness. With the increasing focus on the fairness of models, we use face recognition as an example application of our framework and propose to compactly visualize the fairness analysis of a model via AUC matrices. We analyze the performance of common face recognition models and empirically show that certain subgroups are at a disadvantage when images are perturbed, thereby uncovering trends that were not visible using the model's performance on subgroups without perturbations.

Motivation & Objective

• To address the gap in fairness evaluation for face recognition models under real-world image degradations such as motion blur, exposure, and noise.
• To develop a framework that jointly measures model robustness and subgroup fairness, going beyond traditional summary statistics.
• To visualize subgroup-specific bias under perturbations using AUC matrices for actionable insights into model fairness.
• To identify failure modes where certain demographic subgroups (e.g., young, male, pale skin) are systematically favored or disadvantaged under perturbations.

Proposed method

• Extends VPSA by introducing subgroup-aware perturbation analysis, measuring model performance across protected attributes (e.g., gender, race, facial features) under incremental image degradations.
• Applies controlled perturbations—such as motion blur, saturation shifts, exposure changes, and JPEG compression—across test images.
• Computes Fair SA curves that plot model performance (e.g., verification accuracy) against perturbation levels for each subgroup, enabling visualization of bias trends.
• Uses AUC matrices to compactly represent fairness across multiple attributes and perturbation types, with L1 norms summarizing overall fairness per model.
• Performs row- and column-wise marginalization of AUC matrices to identify the most impactful attributes and perturbations on fairness.
• Employs self-matching and verification tasks to evaluate fairness under identity-agnostic and identity-specific conditions, respectively.

Experimental results

Research questions

• RQ1 How do common face recognition models perform under incremental image perturbations when evaluated for fairness across demographic subgroups?
• RQ2 Which demographic subgroups are systematically favored or disadvantaged under specific types of image degradation?
• RQ3 To what extent does model robustness, as measured by VPSA, correlate with fairness under perturbations?
• RQ4 How can fairness under perturbations be visualized and summarized effectively for model evaluation and debugging?
• RQ5 Can Fair SA detect hidden biases that are absent in standard accuracy metrics but emerge under real-world image degradations?

Key findings

• Even models with 99% accuracy on clean benchmarks show significant subgroup bias under perturbations, such as favoring young individuals under motion blur.
• VGGFace2, while most robust under VPSA, is the least fair model at high perturbation levels, particularly disadvantaging male and pale skin subgroups under exposure and saturation changes.
• FaceNet shows a non-monotonic fairness trend under motion blur, where bias initially increases then decreases, likely due to performance degradation at high blur levels.
• Perturbations that degrade salient facial features—such as blur or compression—most strongly affect subgroups with prominent features (e.g., older individuals with distinct wrinkles or facial hair).
• Speckle noise, motion blur, and saturation shifts are the most detrimental perturbations to fairness across models, as shown by high column-wise L1 norms in AUC matrices.
• AUC matrices effectively reveal hidden fairness trends, such as models favoring black hair under high exposure and heavy makeup under JPEG compression, which are undetectable via standard accuracy metrics.