Light CNN Architecture Enhancement for Different Types Spoofing Attack Detection

The widely acknowledged vulnerability of automatic speaker verification systems (ASV) to various spoofing attacks requires the development of countermeasures robust to unforeseen spoofing trials. In this paper we consider deep learning approach based on Light CNN architecture and its modification for replay attack detection on the base of ASVspoof2017 V2. The efficiency of Light CNN based approaches for replay attacks detection has already been confirmed during ASVspoof2017 (for ASVspoof V1 corpora) and ASVspoof2019 Challenges. We enhanced Light CNN architecture previously considered by the authors via applying angular margin based softmax activation for training robust deep Light CNN classifier. The proposed system achieved Equal Error Rate of 5.5% on the evaluation part of ASVspoof2017 V2. In addition, we also investigated the possibility of unified LCNN-based approach to detect not only replay spoofing attacks but also attacks of logical level, specifically speech synthesis and voice conversion. The experiment results were obtained for microphone part of PHONESPOOF database.