Robustness of audio deepfake detection under audio watermarks

Researchers treat audio watermarks as structured non-adversarial perturbations, build an evaluation framework based on WavMark, test multiple self-supervised, CNN, and GNN audio deepfake detection models, and find that the degree of degradation is highly dependent on the dataset. Watermarks cause significant performance drops on ASVspoof 2021 LA and DF, with limited impact on ASVspoof 2024, FoR, and ITW. The larger the embedding space shift measured by Fréchet Distance, cosine similarity, and L2 distance, the more severe the detection performance drop; the authors therefore argue that benign signal transformations designed for content protection expose previously overlooked robustness weaknesses in current detection systems. The code is open-sourced on GitHub.