Anatomizing Bias in Facial Analysis

Richa Singh, Puspita Majumdar, Surbhi Mittal, Mayank Vatsa

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.

Introduction

As AI moves from the lab to high-stakes environments like criminal sentencing and healthcare, the "black box" of algorithmic bias has become a critical societal issue. Facial analysis systems frequently underperform on specific demographic subgroups, particularly darker-skinned individuals and women. This paper provides a comprehensive systematic review and taxonomy of how bias manifests in facial analysis systems—from face recognition to attribute prediction. Rather than viewing bias as a simple data problem, this work "anatomizes" the entire system pipeline to show how disparities emerge during pre-processing, feature extraction, and final decision-making.

The Anatomy of Bias

A core insight of this research is that bias is a pipeline problem. It is not solely a result of imbalanced training data; rather, it can be introduced or amplified at multiple stages of the machine learning workflow.

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

As shown in the figure above, bias can creep in during:

  • Pre-processing (b): Quality assessment algorithms might reject certain skin tones or facial features.
  • Feature Extraction & Model Training (c): Models may learn latent variables that correlate with protected attributes even when those attributes are not explicitly labeled.
  • Output Prediction (d): Using a single "global" similarity threshold may work well for a majority group but fail for others, necessitating subgroup-specific thresholds.

Mitigation Strategies

The paper categorizes modern mitigation strategies into several algorithmic families:

  • Deep Learning-Based Approaches: Designing novel loss functions, such as "Race adaptive margin loss," to penalize errors on underrepresented groups more heavily.
  • Adversarial Learning: Employing "de-biasing" networks where one component attempts to recognize a face while another tries to "unlearn" protected attributes like gender or race from internal representations.
  • Generative Approaches: Utilizing Generative Adversarial Networks (GANs) to synthesize realistic images of underrepresented groups, thereby balancing training datasets.
  • Information-Theoretic Methods: Applying techniques like mutual information minimization to ensure classification features remain statistically independent of sensitive attributes.

Datasets for Bias Research

The study identifies 14 major databases critical for bias research. Notable examples include RFW (Racial Faces in-the-Wild) for testing racial bias across four subgroups, FairFace for balanced training across seven race subgroups, and MORPH, which facilitates longitudinal studies on the effects of aging. These datasets highlight that while skin tone is a significant factor, facial morphology and broader ethnic features are also essential for understanding performance disparities.

Key Results

A meta-analysis of verification accuracy on the RFW dataset reveals a promising trend in the field.

Verification accuracy trends across different racial subgroups for state-of-the-art models.

While older models (circa 2019-2020) showed significant performance gaps—with African and Asian subgroups often lagging 5-10% behind Caucasians—state-of-the-art models from 2021 (e.g., RamFace, InsideBias) have narrowed this gap significantly. Many now achieve over 95% accuracy across all four major racial subgroups. However, a "fairness-performance trade-off" often persists, where mitigating bias can lead to slight accuracy drops for the majority group.

Challenges and Future Directions

Despite recent progress, several "frontier" problems remain:

  • Intersectional Fairness: Most studies look at race or gender in isolation. Real-world bias is often intersectional (e.g., failing specifically on Black women despite being fair to women and Black people generally).
  • The Impact of Face Masks: The ubiquity of masks introduces new variables that may disproportionately affect recognition accuracy across different demographics.
  • Privacy vs. Fairness: Regulations like GDPR make collecting the sensitive demographic labels needed to fix bias increasingly difficult, creating a "fairness-privacy paradox."
  • Metadata Consistency: Lack of standardized categorization for race and ethnicity across databases complicates cross-study comparisons.

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