This paper investigates the robustness of Wav2Vec2-based audio deepfake detection under real-world telephony degradation conditions, including G.711 codec compression and additive white Gaussian noise (AWGN) at multiple signal-to-noise ratios. Using the ASVspoof 2019 Logical Access dataset, we conduct a detailed per-attack evaluation across 13 unseen attack types under 10 degradation conditions. Our results reveal three key findings: (1) G.711 codec compression significantly improves detection accuracy from 85.89% to 93.28%, contradicting conventional assumptions that compression degrades performance. (2) Noise sensitivity exhibits non-monotonic behavior, with accuracy dropping at 30 dB SNR, recovering above the clean baseline at 20 dB, and degrading sharply at lower SNR levels. (3) Attack-specific analysis reveals heterogeneous vulnerability patterns, including attacks that improve under noise and others that are highly sensitive to specific degradation types. These findings suggest that telephony preprocessing may actively enhance detection performance and highlight the importance of per-attack evaluation for real-world deployment of audio deepfake detection systems.
Sheshi Vardhan (Thu,) studied this question.