Reproduce “WaNet -- Imperceptible Warping-based Backdoor Attack” 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

Anh Nguyen, Anh Tran

Researchers at VinAI Research and Hanoi University of Science and Technology developed WaNet, a novel backdoor attack that uses imperceptible image warping to embed malicious behavior in deep learning models. This method achieves near 100% attack success rates on various image classification tasks while maintaining high clean accuracy and effectively evading both human detection and state-of-the-art machine defenses.

Key insights

  • Subtle geometric image warping can create highly imperceptible backdoor triggers, making them significantly harder for humans to distinguish from original images compared to additive patterns.
  • Training with a "noise mode" is crucial for compelling a network to learn a specific geometric warp, thereby preventing it from relying on pixel-wise artifacts and evading trigger-optimization based defenses.
  • Current backdoor detection strategies, which often target localized patches or distinct spectral signatures, are insufficient against distributed, content-preserving manipulations like warping.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.