Analyzing and Addressing Bias in Face Recognition Using InsightFace Embeddings

AI can exhibit bias in different domains such as healthcare [7], recruiting [8], and insurance [11], impacting disadvantaged populations which can be subject to less accurate predictions. This project focuses on bias in Face Recognition applications. In particular, how the prediction performance of demographic attributes such as age and gender changes across different ethnicities, since errors and biases in this kind of systems can have serious consequences. In order to investigate bias, the UTKFace dataset will be used since besides face images, it contains as well demographic data: age, gen- der and ethnicity. This project uses InsightFace, an open-source library for 2D and 3D deep face analysis, that will be used to create the embeddings of facial images in the UTKFace dataset. On top of the embeddings, classification layers will be added in order to be able to predict age, gender and ethnicity. The model will be trained on a randomly sampled subset of the UTKFace dataset to maintain the original distribution of the dataset. After training, the accuracy will be computed, together with additional metrics such as group-wise classification accuracy, to assess potential biases across different groups. The dataset will be balanced and a weighted loss technique will be implemented, in order to mitigate any existing bias. The model will be trained from scratch under different procedures to assess its impact on the fairness and performance of the predictions.