Reducing Bias in Face Recognition

In this chapter, we estimate the demographic bias in FR (Face Recognition) algorithms and introduce two methods to mitigate the demographic impact on FR performance. The goal of this research is to learn a fair face representation, where the faces of every group could be equally well-represented. Specifically, we explore de-biasing approaches by designing two different network architectures using deep learning. Meanwhile, we evaluate the model’s demographic bias on various datasets to show how much bias is mitigated in our attempt at improving the fairness of face representations extracted from CNNs.