Source-linked AI summary
Face Recognition: Too Bias, or Not Too Bias?
Joseph P Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, Samson Timoner
TL;DR
Facial-recognition systems can exhibit demographic performance gaps because a single global threshold does not fit all subgroup score distributions. Using the balanced BFW dataset, the paper evaluates subgroup-specific thresholds and human perception, finding more balanced performance and a notable overall-performance boost with adaptive thresholds.
Problem
Global facial-recognition thresholds can create skewed errors and performance gaps across demographic subgroups, while imbalanced evaluation sets favor majority demographics.
Method
The paper introduces the gender- and ethnicity-balanced BFW dataset, evaluates state-of-the-art systems by subgroup, applies adaptive subgroup-specific thresholds, and surveys human perception.
Results
Subgroup-specific thresholds mitigate performance gaps and produce a notable boost in overall performance, while human evaluation supports bias in human perception.
Takeaways & Limitations
Selective similarity thresholds provide an easily adoptable way to obtain more balanced facial-recognition performance across subgroups.
Takeaways & Limitations
The subgroup definitions simplify complex race, ethnicity, and gender identities and vary from definitions used by other sources.
Abstract
from arXiv · showhide
We reveal critical insights into problems of bias in state-of-the-art facial recognition (FR) systems using a novel Balanced Faces In the Wild (BFW) dataset: data balanced for gender and ethnic groups. We show variations in the optimal scoring threshold for face-pairs across different subgroups. Thus, the conventional approach of learning a global threshold for all pairs resulting in performance gaps among subgroups. By learning subgroup-specific thresholds, we not only mitigate problems in performance gaps but also show a notable boost in the overall performance. Furthermore, we do a human evaluation to measure the bias in humans, which supports the hypothesis that such a bias exists in human perception. For the BFW database, source code, and more, visit github.com/visionjo/facerec-bias-bfw.
1. Introduction
The paper examines demographic bias in facial-recognition verification, focusing on how global similarity thresholds produce uneven subgroup errors. It introduces balanced evaluation and subgroup-specific thresholds to measure and mitigate these disparities.
- Problem: Global thresholds can produce skewed facial-recognition errors because similarity-score distributions differ across demographic subgroups.Evaluation sets dominated by majority demographics can further favor common traits in reported performance.
- Scope: The authors define eight subgroups using facial characteristics associated with gender and ethnicity, while acknowledging that these categories simplify complex identities.The paper notes that race, ethnicity, and gender definitions vary across sources.
- Approach: BFW is a balanced benchmark designed to report verification performance separately across demographic subgroups.The dataset addresses evaluation imbalance while enabling per-subgroup ratings.
- Approach: Subgroup-specific adaptive thresholds are proposed to reduce performance disparities and increase accuracy relative to using one global threshold.The thresholds vary with detected facial attributes, specifically gender and ethnicity.
- Human evaluation: The study also evaluates human perception to investigate whether bias occurs in human judgments of faces.The human survey is presented as a third contribution alongside dataset construction and threshold analysis.
2. Background Information
The background situates facial-recognition bias within broader machine-learning fairness research and prior efforts to address demographic imbalance. It distinguishes BFW through balanced gender-ethnicity subgroups and connects machine bias with documented human perceptual bias.
- Prior work: Prior facial-recognition bias research has examined augmentation, one-shot learning, fairness, privacy, domain adaptation, demographic attributes, data exploration, and commercial systems.These approaches reflect different sources and formulations of bias in facial-recognition systems.
- BFW: BFW organizes database statistics and pair counts by ethnicity and then gender, supporting balanced subgroup evaluation.Table 1 describes subgroup definitions, database statistics, and pair partitions.
- Datasets: Existing paired-face datasets studied demographic or age differences, whereas BFW adds Indian subjects, more faces, and gender splits within subgroups.The paper positions BFW as related to prior datasets built by sampling existing databases but expanded for its target analysis.
- Human bias: Bias is not unique to machine learning: prior work reports human differences in recognizing faces across race, gender, and age.The paper cites human performance below random levels for distinguishing some East Asian face categories and reports a similar below-random average in its own evaluation.
3. The BFW Benchmark and Dataset
BFW is a balanced facial-verification benchmark designed to compare performance across identity, gender, and ethnicity subgroups. It uses controlled sampling, validation, pair construction, and subgroup analyses to expose differences in face-score behavior.
- The data: BFW balances identities, gender, and ethnicity, while providing more samples per subject and subgroup than DemogPairs using the single VGG2 resource.Unlike RFW, which emphasizes race distribution and domain adaptation, BFW emphasizes balanced identity and subgroup evaluation.
- The data: Candidates were sampled from VGG2 and refined with ethnicity and gender processing before face detection, alignment, and visual validation.The pipeline stored detections and landmarks, aligned faces, and used Sphereface encodings to help remove false-positive detections.
- Evaluation analyses: Across subgroup analyses, DET curves compare gender, ethnicity, and combined subgroup trade-offs, while score distributions and confusion matrices examine subgroup-specific error patterns.The human evaluation focuses on White American and Chinese participants and look-alike face pairs, reducing variability relative to broader groups.
- The data: BFW constructs balanced five-fold protocols with 25 faces per subgroup subject, no identity overlap across folds, and matched within- and cross-subgroup negative pairs.Subjects were assigned to folds by pair counts to preserve balance, with additional negative pairs sampled across subgroups.
- Problem formulation: Facial verification encodes images into feature vectors and classifies pairs as genuine or impostor using a similarity or distance score.The benchmark uses cosine similarity to measure feature closeness between paired faces.
4. Results and Analysis
Experiments show subgroup-dependent performance differences across face-recognition models, score distributions, and fixed-threshold operating points. Subgroup-specific thresholds reduce these disparities, while human evaluation also reveals subgroup-linked recognition patterns.
- Subgroup-specific thresholds reduce percent differences from the intended FPR to near zero, balancing the remaining small offsets across subgroups.Figure 6 compares global-threshold behavior with category-based thresholds at intended FPR settings.
- DET analysis finds that M performs better than F, while AF and IF perform worst across subgroup comparisons.
- Genuine-pair score distributions vary across subgroups in location and spread, whereas imposter scores remain similarly centered near zero.
- The same approximate subgroup ordering appears across VGG16, ResNet50, and SENet50 backbones, with WM best and AF worst.Figure 7 reports DET curves for VGG2 models using three different CNN backbones.
- Fixed thresholds produce different operating points across subgroups, with potential false-positive differences of about double those reported from aggregate performance.
- Human assessment shows that people are generally most accurate at recognizing faces from their own subgroup, with some exceptions.The quantitative human-assessment table identifies WF pairs as the most correctly matched.