Defining Bias in Face Recognition
In face recognition, bias refers to systematic differences in accuracy across demographic groups, not random variation, but consistent patterns where certain groups are served worse than others. Bias manifests as higher false non-match rates (failing to match the same person), higher false positive rates (matching different people), or lower quality similarity scores for particular groups.
Bias can also be more subtle: a system might achieve similar verification accuracy across groups but produce lower-quality ranked match lists for some users, or have confidence calibration that is accurate on average but systematically miscalibrated for specific subgroups.
Sources of Bias
The dominant source of bias in current systems is training data imbalance. Most large face recognition datasets were assembled from internet data, which overrepresents certain demographics, particularly younger, lighter-skinned, male individuals from certain geographic regions. A model trained on such data develops more precise representations for well-represented groups.
Annotation bias is a secondary source: human labellers who annotate training data make more errors on faces from groups they are less familiar with (the own-race effect), introducing label noise that disproportionately affects underrepresented groups. Evaluation bias is also a concern: models are often benchmarked on datasets that do not represent deployment populations, masking disparities that emerge in practice.
Technical Approaches to Mitigation
Several technical approaches address training data bias. Rebalancing: adjusting the training sampling procedure to ensure balanced representation across demographic groups, so the loss function receives comparable signal from all groups. Adversarial debiasing: training an auxiliary network to predict demographic attributes and penalising the main model if the embedding is predictive of those attributes. Fairness-aware loss functions: directly optimising for equal accuracy across groups alongside overall performance.
None of these approaches eliminates bias entirely; they trade off demographic equity against overall accuracy in ways that require principled decisions about what constitutes acceptable performance. Transparency about evaluated performance across groups, and ongoing monitoring after deployment, are minimum standards for responsible systems.