Gender Discrimination Versus Speaker Identification Through Privacy-Aware Adversarial Feature Extraction

In this paper we propose a deep neural network based feature extraction scheme in order to improve the privacy vs. utility trade-off in speaker classification tasks. In the investigated scenario we aim to maximize the classifier's gender discrimination performance without providing access to more specific speaker identity information. Given these closely related objectives it can be expected that features extracted for gender classification will also carry some speaker-discriminative information. Our privacy enhancement approach consists of employing an adversarial feature extraction model, which aims to maximize gender classification accuracy (utility) while encouraging poor speaker identification results (privacy). The proposed model's loss function and budget scaling factor for controlling the balance between the levels of privacy and utility are analyzed. It is experimentally shown that this approach noticeably reduces the aforementioned privacy risks without provoking a significant utility loss, and that the effects are also generalizable to previously unseen speech sources.