RoGA: Towards Generalizable Deepfake Detection through Robust Gradient Alignment

TL;DR

RoGA tackles cross-domain deepfake detection by aligning gradient updates across domains through parameter perturbations, guiding optimization toward flatter, more robust minima without extra regularization. The method combines a perturbation-based robustness objective with a domain-aware gradient alignment term, extending ideas from sharpness-aware optimization to multi-domain settings. Empirical results across FF++ variants, Celeb-DF, and DFDC show RoGA surpasses state-of-the-art domain generalization methods, with strong ablations validating the contributions of both objectives and backbone independence. The approach yields improved cross-domain detection performance and interpretable representations, suggesting practical impact for deploying deepfake detectors in diverse real-world scenarios.

Abstract

Recent advancements in domain generalization for deepfake detection have attracted significant attention, with previous methods often incorporating additional modules to prevent overfitting to domain-specific patterns. However, such regularization can hinder the optimization of the empirical risk minimization (ERM) objective, ultimately degrading model performance. In this paper, we propose a novel learning objective that aligns generalization gradient updates with ERM gradient updates. The key innovation is the application of perturbations to model parameters, aligning the ascending points across domains, which specifically enhances the robustness of deepfake detection models to domain shifts. This approach effectively preserves domain-invariant features while managing domain-specific characteristics, without introducing additional regularization. Experimental results on multiple challenging deepfake detection datasets demonstrate that our gradient alignment strategy outperforms state-of-the-art domain generalization techniques, confirming the efficacy of our method. The code is available at https://github.com/Lynn0925/RoGA.

Figures (4)

• Figure 1: Comparison among previous cross-domain deepfake detection based on domain adaption such as data/feature augmentation. Our approach in (b) aligns the perturbed gradients, enhancing robustness and generalization.
• Figure 2: The left part illustrates the framework of RoGA. The right part demonstrates the basic idea of RoGA, where it aligns the gradients with different perturbations during training, hence reducing the risk of overfitting to specific domain patterns.
• Figure 3: The GradCAM visualizationsGRADCAM comparing SRM super-resolution baseline and our RoGA, across four forgery types on FF++(c23).
• Figure 4: t-SNEtsne visualization of latent space w and w/o our RoGA when the model is trained on FF++(c23).