HANUI: Harnessing Distributional Discrepancies for Singing Voice Deepfake Detection

In this work, we propose a novel framework for deepfake detection in singing voices, based on explicitly capturing the distributional differences between bona fide and spoofed signals. While prior approaches employing self-supervised models or graph neural networks have shown promising results, they remain vulnerable when tested on unseen singers, musical styles, and languages. Motivated by anomaly detection, our method integrates an autoencoder with a GAN-based architecture to exploit probabilistic distributional discrepancies between ground-truth and reconstructed signals. In particular, we leverage a discriminator to extract informative feature maps that highlight distinctive characteristics of bona fide and spoofed samples, which are subsequently utilized by a detector to perform classification. Experimental results demonstrate that our proposed framework not only improves detection accuracy over recent methods, but also achieves substantial relative reductions in error rates, confirming its robustness and generalizability under challenging unseen conditions.