Experimental Case Study of Self-Supervised Learning for Voice Spoofing Detection

This study aims to improve the performance of voice spoofing attack detection through self-supervised pre-training. Supervised learning needs appropriate input variables and corresponding labels for constructing the machine learning models that are to be applied. It is necessary to secure a large number of labeled datasets to improve the performance of supervised learning processes. However, labeling requires substantial inputs of time and effort. One of the methods for managing this requirement is self-supervised learning, which uses pseudo-labeling without the necessity for substantial human input. This study experimented with contrastive learning, a well-performing self-supervised learning approach, to construct a voice spoofing detection model. We applied MoCo's dynamic dictionary, SimCLR's symmetric loss, and COLA's bilinear similarity in our contrastive learning framework. Our model was trained using VoxCeleb data and voice data extracted from YouTube videos. Our self-supervised model improved the performance of the baseline model from 6.93% to 5.26% for a logical access (LA) scenario and improved the performance of the baseline model from 0.60% to 0.40% for a physical access (PA) scenario. In the case of PA, the best performance was achieved when random crop augmentation was applied, and in the case of LA, the best performance was obtained when random crop and random shifting augmentations were considered.