Audio Deepfake Detection Under Post-Processing Attack

Generalizable audio deepfake detection is a challenging task. Simple post-processing attacks like background noise or impulse response can significantly affect the performance of the detectors. We analysed the effects of 13 post-processing attacks on two detectors, one with a SSL (Self-Supervised Learning)-based front-end (Wav2Vec 2.0) the other using SincNet for feature extraction. Both detectors showed significant performance degradation when applying the post-processing attacks. For instance, we calculated an EER of 0.73% on the original data of the in-the-wild dataset using the SSL-based detector. The performance dropped to 4.37% after applying impulse response. To find the most effective attacks, we analysed the effects of post-processing on their signal quality using UTMOS. Additionally, we explored retraining strategies, improving the overall performance of our detectors by an EER of 0.22% and 0.33%.