The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection

The article discusses the vulnerabilities introduced by provenance watermarking in synthetic speech detection, highlighting how models like Chatterbox and techniques such as AudioSeal can lead to significant detection failures. It identifies three key issues: generalization degradation, strip-to-evade, and mark-to-frame, with empirical results showing that a watermark-trained detector's Equal Error Rate can increase from 16% to 75%. The authors propose retraining detectors with watermarks present in both synthetic and human speech to mitigate these issues and release the WASP corpus for further research.