Similarity Learning Based Adversarial Training for Censoring Representatoins

Abstract

Deep neural networks (DNN) continuously demonstrates excellent performance in various application domains. However, how to control the representations is critical issues to use DNN in real-world scenarios. Notably, control the invariance of the representations is essential to incorporate social constraints, such as privacy-protection and fairness. This paper proposes a novel way to control the representations learned by DNN, called similarity confusion training. Empirical validations on a task of learning anonymous representations from the data of wearables show that the proposed method successfully remove unwanted information with less performance degradation compared to the existing methods.