Measuring the robustness of audio deepfake detection under real-world distortions

The researchers systematically evaluated the robustness of 10 audio deepfake detection models under 18 common distortions, which fall into three categories: noise perturbation, audio modification and compression; the tested models cover traditional deep learning models and speech foundation models. Results show that most models are relatively robust to noise, but performance drops markedly under audio modification and compression, with the impact of neural codecs especially prominent; speech foundation models consistently outperform traditional models in most distortion scenarios, which the authors say may stem from their large-scale pretraining on diverse audio data; increasing model scale can improve robustness but with diminishing returns; targeted data augmentation during training or speech enhancement at inference can improve robustness to unseen distortions. The paper emphasizes that deepfake detection deployed in practice should be evaluated under diverse real-world distortions (arXiv:2503.17577).