Anatomizing Bias in Facial Analysis

Richa Singh, Puspita Majumdar, Surbhi Mittal, Mayank Vatsa

Introduction

Facial analysis technologies, including face recognition, gender classification, and attribute prediction, are increasingly deployed in high-stakes sectors such as law enforcement and healthcare. Despite their utility, these systems frequently demonstrate "demographic bias," exhibiting significantly lower performance for specific subgroups based on gender, ethnicity, or skin tone. This paper provides a comprehensive survey and taxonomy of bias in facial analysis, moving beyond general observations to identify the specific stages where discrimination enters the AI pipeline.

The Facial Analysis Pipeline and Bias Introduction

Bias is not a monolithic issue but rather a cumulative effect of decisions made throughout the machine learning lifecycle. The authors anatomize the pipeline into distinct stages—data collection, pre-processing, model training, and final prediction—each offering opportunities for both the introduction and mitigation of bias.

The facial analysis pipeline identifying stages (a-d) where bias can be introduced or mitigated.

Mitigation Strategies

The paper categorizes modern mitigation techniques into several high-level approaches:

  • Deep Learning-Based Mitigation: Implementing novel loss functions, such as "Sensitive Loss" or "Race Adaptive Margin Loss," which penalize models for disparate error rates across groups.
  • Adversarial Learning: Utilizing "de-biasing" networks where a model is trained for recognition while a competing adversarial component ensures internal features do not encode protected attributes like race or gender.
  • Generative Approaches: Using GANs to synthesize images of under-represented groups, creating balanced virtual datasets to overcome the privacy and logistical hurdles of real-world data collection.
  • Fair Score Normalization: Adjusting final output scores based on group fairness metrics without requiring a full model retraining.

Benchmarking with Diverse Datasets

A critical component of bias research is the availability of diverse and well-annotated datasets. The paper references 14 key databases used for bias research, including:

  • MORPH: A large-scale dataset for studying age, gender, and race.
  • Racial Faces in-the-Wild (RFW): Designed specifically to test racial bias in unconstrained environments.
  • FairFace: A benchmark balanced across seven race groups to prevent over-fitting to majority demographics.
  • IJB-C: A dataset incorporating skin-tone information based on the Fitzpatrick scale.

A meta-analysis of verification accuracy on the RFW dataset reveals both progress and persistent challenges. While newer algorithms (circa 2021) have successfully narrowed the accuracy gap between Caucasian, Indian, Asian, and African subgroups compared to earlier models, a performance hierarchy remains.

Verification accuracy across different racial subgroups for state-of-the-art models (2020-2021).

As shown in the figure above, Caucasian groups generally experience the highest accuracy, while performance for African and Asian groups remains more volatile across different algorithmic implementations.

Key Insights and Open Challenges

Beyond demographic labels, research indicates that bias is deeply embedded in system design and pre-processing:

  • The "Global Threshold" Trap: Using a single decision threshold for all groups often favors majority demographics. Optimal thresholds vary by subgroup.
  • Pre-processing Bias: Tools like "face quality assessment" often fail more frequently for individuals with darker skin or young children, filtering them out before the core model even processes the data.
  • Latent Encoding: Models often "discover" skin color in latent representations even when not explicitly trained on it, with high latent values frequently correlating with lighter skin tones.

Future Directions

The authors identify several frontiers for future research:

  • Intersectional Fairness: Moving beyond single-attribute mitigation to address bias against specific intersections, such as older, darker-skinned females.
  • Fairness Under Occlusion: Investigating how standard occlusions, like face masks, impact demographic fairness.
  • The Fairness-Performance Trade-off: Developing methods to reduce bias without incurring a significant "accuracy tax" on overall system performance.

Summary updated

Abstract

Existing facial analysis systems have been shown to yield biased results against certain demographic subgroups. Due to its impact on society, it has become imperative to ensure that these systems do not discriminate based on gender, identity, or skin tone of individuals. This has led to research in the identification and mitigation of bias in AI systems. In this paper, we encapsulate bias detection/estimation and mitigation algorithms for facial analysis. Our main contributions include a systematic review of algorithms proposed for understanding bias, along with a taxonomy and extensive overview of existing bias mitigation algorithms. We also discuss open challenges in the field of biased facial analysis.
BibTeX
@inproceedings{singh2022anatomizing,
  author = {Richa Singh and Puspita Majumdar and Surbhi Mittal and Mayank Vatsa},
  title = {{Anatomizing Bias in Facial Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21500},
}
Figure