A Novel Approach Towards Generalization of Countermeasure for Spoofing Attack on ASV Systems

Automatic speaker verification (ASV) system is a voice-based biometric authentication system susceptible to attacks from spoofed/fake utterances generated using speech synthesis, voice conversion and record-replay to gain fraudulent access through the system. Development of a generalized countermeasure to detect all such attacks is a major concern to the research community today. This paper presents a new feature named filter-based cepstral coefficient (FBCC) used in the front-end processing of the countermeasure. FBCC is cepstral coefficients of discrete cosine transformed log compressed energy variations of presented utterances. The energy variation patterns (EVPs) are captured by applying statistical filters to the power spectral density of utterance presented to the ASV system. Cepstral coefficients retrieved from EVP tend to capture artefacts that well discriminates spoofed from bonafide utterances. The robustness of FBCC in projecting the traces of spoof attack in the utterances is evident through the empirical results obtained when tested on ASVspoof2019 corpus with the conventional Gaussian mixture model classifier. The performance of the countermeasure using FBCC is consistently good throughout all the presented attacks, thus achieving generalization. This FBCC-based countermeasure performed equally good under both logical access (LA) and physical access (PA) conditions. The best performance of FBCC-based countermeasure, when tested with different filters, using t-DCF (tandem detection cost function) metric is 0.00 and 0.16 for development and evaluation data, respectively, under LA condition and is 0.18 and 0.25 for development and evaluation data, respectively, under PA condition.