Suppressing Gradient Conflict for Generalizable Deepfake Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deepfake detection models require generalization to evolving manipulation techniques.
• Augmenting training data with online synthesized forgeries is a promising strategy.
• Jointly training on real and fake data can surprisingly degrade performance due to gradient conflicts.

Purpose of the Study:

• To investigate the cause of performance degradation when training deepfake detection models on both real and synthesized data.
• To propose a novel framework that mitigates gradient conflicts for improved deepfake detection.
• To enhance both in-domain accuracy and cross-domain generalization of deepfake detection models.

Main Methods:

• Proposed the Conflict-Suppressed Deepfake Detection (CS-DFD) framework.
• Introduced an Update Vector Search (UVS) module to find alternative gradient updates.
• Developed a Conflict Gradient Reduction (CGR) module with a novel Conflict Descent Loss to align gradients.

Main Results:

• The CS-DFD framework effectively mitigates gradient conflicts between real and synthesized data.
• UVS module optimizes updates to maximize simultaneous loss reduction for both data types.
• CGR module enforces a low-conflict feature embedding space, improving representation learning.
• CS-DFD achieved state-of-the-art performance on deepfake detection benchmarks.

Conclusions:

• Gradient conflicts are a key factor limiting deepfake detection model generalization.
• The proposed CS-DFD framework successfully addresses these conflicts.
• CS-DFD demonstrates superior performance in both accuracy and generalization capabilities for deepfake detection.