Researchers Bypass Deepfake Detection With Replay Attacks

Researchers from several universities and Resemble AI demonstrated that replay attacks (playing synthetic audio and re-recording it through real speakers/microphones, sometimes with background noise) significantly reduce the effectiveness of audio deepfake detection models; they released a 132.5-hour ReplayDF dataset for noncommercial use and found top model error rates increased (e.g., W2V2-AASIST EER rose from 4.7% to 18.2%). Retraining with room impulse responses improved resilience but did not fully mitigate the vulnerability; authors advise defenders to rely on content verification and out-of-band confirmation to counter vishing using voice cloning.