Exploring Racial Bias in Deep Face Recognition Models

Deep convolutional neural networks (DCNNs) have played a significant role in the advancement of face recognition algorithms, demonstrating exceptional generalization capabilities on very large datasets. In the case of deep face recognition systems, models trained on racially unbalanced datasets-particularly those skewed towards majority Caucasian demographics-encounter difficulties in accurately recognizing faces from underrepresented groups. As such, this paper presents an evaluation of several state-of-the-art deep face recognition models on racially balanced datasets. The results reveal substantial performance disparities between the models on widely used benchmarks like Labeled Faces in the Wild (LFW) and unbiased datasets of comparable sizes. Of the models evaluated, FaceNet emerges as the model with the highest proficiency in generalizing across racial minorities. Emphasizing the significance of comprehensive performance reporting, we additionally advocate for metrics beyond accuracy, such as the F1 score. These additional metrics allow for derivation of new thresholds that enhance the evaluations and effectively optimize model performance across the balanced data.