Reproduce “Creating Blank Canvas Against AI-enabled Image Forgery” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Song, Ziyuan Luo, Renjie Wan
Researchers at Hong Kong Baptist University developed a proactive, training-free method to detect AI-enabled image forgery by transforming original images into a "blank canvas" using frequency-aware adversarial perturbations. This approach leverages an off-the-shelf Segment Anything Model (SAM) to localize tampering, achieving F1-scores exceeding 95% and IoU around 90% against diverse AI-generated content manipulations.
Key insights
- Tampering becomes inherently more conspicuous when a vision model is first "blinded" to the original image content, turning anomaly detection into an effective tamper localization mechanism.
- Precisely targeting high-frequency components through frequency-aware adversarial perturbations is crucial for thoroughly disrupting the perception of powerful foundation models like SAM, which are robust to naive attacks.
- Leveraging off-the-shelf vision foundation models for security applications without task-specific retraining offers a generalizable and resource-efficient approach to counter evolving AI-enabled threats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.