Towards JPEG-Resistant Image Forgery Detection and Localization Via Self-Supervised Domain Adaptation.

Area of Science:

• Digital Image Forensics
• Computer Vision
• Machine Learning

Background:

• Image forgery is a growing concern due to widespread editing tools.
• Current forgery detection methods struggle with JPEG compression common in online sharing.
• A domain adaptation approach is needed to address the performance gap caused by compression.

Purpose of the Study:

• To develop a JPEG-resistant method for image forgery detection and localization.
• To improve the robustness of forgery detection against various JPEG compression levels.
• To mitigate the domain shift between uncompressed and compressed images.

Main Methods:

• A self-supervised domain adaptation network combining a Siamese backbone and a Compression Approximation Network (ComNet).
• ComNet approximates JPEG compression via self-supervised learning to create general JPEG-agent images.
• Domain adaptation training alleviates discrepancies between uncompressed and JPEG-agent image domains.

Main Results:

• The proposed method demonstrates superior or competitive performance compared to state-of-the-art techniques.
• Effective forgery detection and localization were achieved, particularly under unknown Quantization Factors (QFs) of JPEG compression.
• The approach shows significant improvements in handling JPEG-compressed forged images.

Conclusions:

• The developed network effectively addresses the challenge of JPEG compression in image forgery detection.
• The self-supervised domain adaptation strategy enhances robustness against common image compression artifacts.
• This method offers a promising solution for real-world scenarios involving online image manipulation.