RoGA: Towards Generalizable Deepfake Detection through Robust Gradient Alignment

2025-05-27

AI Summary

Existing deepfake detection methods suffer from source-domain overfitting during cross-domain generalization, while conventional regularization strategies often degrade empirical risk minimization (ERM) performance. Method: We propose a robust gradient alignment learning framework that requires no additional regularization. By applying adversarial parameter perturbations, it explicitly aligns the ERM gradient directions across domains, thereby encouraging domain-invariant feature learning. Contribution/Results: Our key innovation is the first direct alignment of ERM gradients with generalization-aware gradients during optimization—eliminating the need for auxiliary regularizers that interfere with the primary task. The method operates within an end-to-end differentiable framework and supports joint multi-domain gradient constraints. It achieves significant improvements over state-of-the-art domain generalization methods on multiple mainstream deepfake detection benchmarks. Code is publicly available.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

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.

Problem

• Enhances deepfake detection robustness to domain shifts
• Aligns generalization gradient updates with ERM objectives
• Preserves domain-invariant features without extra regularization

Innovation

• Aligns generalization and ERM gradient updates
• Applies perturbations to model parameters
• Enhances robustness to domain shifts