Active Adversarial Noise Suppression for Image Forgery Localization

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Deep learning advances image forgery localization but models are vulnerable to adversarial attacks.
• Imperceptible noise in adversarial attacks significantly misleads current forgery localization models.

Purpose of the Study:

• To develop a defense mechanism against adversarial attacks for image forgery localization.
• To enhance the robustness of deep learning models in detecting manipulated images.

Main Methods:

• Introduced an Adversarial Noise Suppression Module (ANSM) generating defensive perturbations.
• Implemented a two-stage training strategy: Forgery-relevant Features Alignment (FFA) and Mask-guided Refinement (MgR).
• FFA minimizes distributional discrepancies using Kullback-Leibler divergence; MgR uses dual-mask constraints for refinement.

Main Results:

• The proposed method significantly restores forgery localization performance on adversarial images.
• Performance on original forged images remains largely unaffected when using the ANSM.
• Extensive experiments confirm effectiveness across various attack algorithms.

Conclusions:

• The developed Adversarial Noise Suppression Module (ANSM) provides effective defense against adversarial attacks in image forgery localization.
• This work represents the first reported adversarial defense specifically for image forgery localization tasks.
• The source code and anti-forensics dataset are released for further research.