Improving Generalization in Deepfake Detection With Feature-Space Adversarial Consistency
IEEE Signal Processing Letters · 2026 · 33 · 2295-2299
DOI: 10.1109/lsp.2026.3698377 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Pratibha Singh und 1 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Gefördert von: U.S. Department of Commerce.
Abstract (Originalsprache)
Deepfake detectors often exhibit limited generalization to unseen datasets due to their tendency to overfit dataset-specific cues, such as manipulation-dependent artifacts present in the training data. Although various strategies have been proposed to reduce this dependence, including architectural modifications, data augmentation, and regularization techniques, cross-domain performance degradation remains a challenging problem. In this letter, we propose Feature-Space Adversarial Consistency (FSAC), a lightweight training-time regularization framework that improves cross-dataset generalization by enforcing prediction stability under worst-case perturbations in a fused feature space. The FSAC generates adversarial perturbations in feature space using gradient-based updates and enforces prediction consistency between clean and perturbed features through a Kullback-Leibler (KL) divergence loss, without fine-tuning the frozen backbone. Extensive cross-dataset experiments on multiple benchmarks demonstrate that FSAC consistently outperforms state-of-the-art methods while introducing no additional inference-time overhead.
Förderung
- U.S. Department of Commerce
Zitieren
APA
Singh, P., & Pankajakshan, V. (2026). Improving Generalization in Deepfake Detection With Feature-Space Adversarial Consistency. IEEE Signal Processing Letters, 33, 2295-2299. https://doi.org/10.1109/lsp.2026.3698377
Vancouver
Singh P, Pankajakshan V. Improving Generalization in Deepfake Detection With Feature-Space Adversarial Consistency. IEEE Signal Processing Letters. 2026;33:2295-2299. doi:10.1109/lsp.2026.3698377
BibTeX
@article{Singh2026,
author = {Singh, Pratibha and Pankajakshan, Vinod},
title = {Improving Generalization in Deepfake Detection With Feature-Space Adversarial Consistency},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2295--2299},
doi = {10.1109/lsp.2026.3698377}
}
EUVIMED