Exploring Racial Bias within Face Recognition via per-subject Adversarially-Enabled Data Augmentation

Whilst face recognition applications are becoming increasingly prevalent\nwithin our daily lives, leading approaches in the field still suffer from\nperformance bias to the detriment of some racial profiles within society. In\nthis study, we propose a novel adversarial derived data augmentation\nmethodology that aims to enable dataset balance at a per-subject level via the\nuse of image-to-image transformation for the transfer of sensitive racial\ncharacteristic facial features. Our aim is to automatically construct a\nsynthesised dataset by transforming facial images across varying racial\ndomains, while still preserving identity-related features, such that racially\ndependant features subsequently become irrelevant within the determination of\nsubject identity. We construct our experiments on three significant face\nrecognition variants: Softmax, CosFace and ArcFace loss over a common\nconvolutional neural network backbone. In a side-by-side comparison, we show\nthe positive impact our proposed technique can have on the recognition\nperformance for (racial) minority groups within an originally imbalanced\ntraining dataset by reducing the pre-race variance in performance.\n