From Sharpness to Better Generalization for Speech Deepfake Detection

Publication
arXiv preprint arXiv:2506.11532

This paper investigates how sharpness-aware optimization techniques can enhance generalization in speech deepfake detection models. The authors analyze the link between model flatness, robustness, and generalization, and propose methods to mitigate overfitting to specific spoofing conditions while maintaining strong cross-dataset performance.