Dominant Tangent Residual Suppression for Robust One-Class Audio Deepfake Detection Against Unseen Attacks
IEEE Access · 2026 · 14 · 163718-163733
DOI: 10.1109/access.2026.3742554 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Seung-Bin Kim und 2 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Supreme Prosecutors’ Office(SPO) grant funded by the Korea governmen.
Abstract (Originalsprache)
Audio deepfake detection increasingly adopts one-class learning to model bona fide speech compactly around a centroid, with OC-Softmax as a representative method. However, OC-Softmax optimizes only the scalar cosine similarity between an embedding and the centroid, reducing the magnitude of the tangent residual while leaving its direction unconstrained. As a result, bona fide residuals may concentrate along a few dominant tangent directions, producing an anisotropic expansion that widens the acceptance region and increases overlap with high-scoring spoof samples. We propose Dominant Tangent Residual Suppression (DTRS), a direction-aware regularizer that estimates the dominant tangent residual subspace from the bona fide residual covariance during training and penalizes the residual energy projected onto this subspace. Operating only during training, DTRS complements the score-level compression of OC-Softmax at the direction level, while leaving the scoring function and the inference cost identical to the baseline. Across five backbones and five benchmarks, DTRS consistently improves the average cross-benchmark equal error rate (EER) and minimum detection cost over OC-Softmax for every backbone, reducing the average EER by 11.1% relative in the best configuration, with the largest single-benchmark gain of 13.4% on the out-of-domain In-the-Wild benchmark. Residual diagnostics, ablation studies, and residual-space visualizations further show that both the low-scoring bona fide tail and high-scoring spoof exhibit a high parallel residual ratio within the dominant subspace, and that DTRS suppresses this directional alignment, widening the separation between the two distributions and improving robustness to the evaluated unseen and real-world attacks.
Förderung
- Supreme Prosecutors’ Office(SPO) grant funded by the Korea governmen
Zitieren
APA
Kim, S., Kim, K., & Yu, H. (2026). Dominant Tangent Residual Suppression for Robust One-Class Audio Deepfake Detection Against Unseen Attacks. IEEE Access, 14, 163718-163733. https://doi.org/10.1109/access.2026.3742554
Vancouver
Kim S, Kim K, Yu H. Dominant Tangent Residual Suppression for Robust One-Class Audio Deepfake Detection Against Unseen Attacks. IEEE Access. 2026;14:163718-163733. doi:10.1109/access.2026.3742554
BibTeX
@article{Kim2026,
author = {Kim, Seung-Bin and Kim, Kyung-Wha and Yu, Ha-Jin},
title = {Dominant Tangent Residual Suppression for Robust One-Class Audio Deepfake Detection Against Unseen Attacks},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {163718--163733},
doi = {10.1109/access.2026.3742554}
}
EUVIMED