Face Recognition Fairness Assessment based on Data Augmentation: An Empirical Study

Abstract

Deep learning models are affected by the training data when classifying, leading to discrimination in prediction output or disparity in prediction quality. We need to test the model adequately using a large amount of data. However, data for certain combinations of attributes occur less frequently in reality and are more difficult to obtain. Data augmentation is one of the methods to alleviate this problem. In this paper, we conduct a preliminary study on whether changes in these features(hair, glasses, bangs, etc.) could affect classification accuracy. This study provides some conclusions, (1) there is a fairness problem in the depth model (2) the fairness of the model can be well tested by auamentation against Image attributes.