Towards generalizable face forgery detection via mitigating spurious correlation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Security

Background:

• Face forgery techniques, particularly deepfakes, pose significant threats to information security and personal privacy.
• Current deep learning-based detection methods perform well within their training domain but struggle with cross-domain generalization.
• Spurious correlations between irrelevant features and forgery labels are identified as a key limitation in existing detectors.

Purpose of the Study:

• To develop a novel method that enhances the generalization capability of face forgery detection models.
• To address the challenge of detecting unknown forgeries and improving cross-domain performance.
• To mitigate the impact of spurious correlations in deepfake detection.

Main Methods:

• Proposed the Feature Independence Constrainer (FIC) to map features into a Reproducing Kernel Hilbert Space and iteratively optimize their covariance matrix, ensuring feature independence.
• Incorporated fine-grained high-frequency components to guide the model towards learning genuine forgery artifacts.
• Introduced a Feature Alignment Module (FAM) to capture higher-order spatial-frequency dependencies for richer trace extraction.

Main Results:

• The proposed method demonstrates competitive performance on multiple face forgery benchmark datasets.
• Quantitative and qualitative experiments confirm the effectiveness of FIC in alleviating spurious correlations.
• The approach outperforms existing state-of-the-art methods in cross-domain face forgery detection scenarios.

Conclusions:

• The Feature Independence Constrainer (FIC) effectively enhances the generalization of deepfake detection models by minimizing spurious correlations.
• The integration of high-frequency components and the Feature Alignment Module contribute to more robust and comprehensive forgery detection.
• This research offers a promising direction for building more reliable face forgery detection systems resilient to domain shifts.