Patch-Discontinuity Mining for Generalized Deepfake Detection

2025-12-26

AI Summary

Existing deepfake detection methods suffer from severe generalization failure against unseen forgery patterns, struggling to address increasingly realistic cross-domain and cross-manipulation threats. To address this, we propose a lightweight generalizable detection framework comprising three key components: (1) a deepfake-specific representation learning mechanism that explicitly models forensic traces; (2) a parameter-free feature-space redistribution strategy coupled with classification-agnostic feature enhancement to improve cross-domain robustness; and (3) synergistic integration of large-vision-model transfer, local patch discontinuity modeling, and parameter-free data augmentation. With only 0.28M trainable parameters, our method achieves state-of-the-art generalization performance under both cross-domain and cross-manipulation evaluation protocols, significantly mitigating the sharp performance degradation commonly observed in unknown forgery scenarios.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: Safety and Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

The rapid advancement of generative artificial intelligence has enabled the creation of highly realistic fake facial images, posing serious threats to personal privacy and the integrity of online information. Existing deepfake detection methods often rely on handcrafted forensic cues and complex architectures, achieving strong performance in intra-domain settings but suffering significant degradation when confronted with unseen forgery patterns. In this paper, we propose GenDF, a simple yet effective framework that transfers a powerful large-scale vision model to the deepfake detection task with a compact and neat network design. GenDF incorporates deepfake-specific representation learning to capture discriminative patterns between real and fake facial images, feature space redistribution to mitigate distribution mismatch, and a classification-invariant feature augmentation strategy to enhance generalization without introducing additional trainable parameters. Extensive experiments demonstrate that GenDF achieves state-of-the-art generalization performance in cross-domain and cross-manipulation settings while requiring only 0.28M trainable parameters, validating the effectiveness and efficiency of the proposed framework.

Problem

• Detects deepfake images using compact network design
• Improves generalization across unseen forgery patterns
• Reduces trainable parameters while maintaining high performance

Innovation

• Utilizes deepfake-specific representation learning for discriminative patterns
• Implements feature space redistribution to address distribution mismatch
• Applies classification-invariant feature augmentation for enhanced generalization