Reproduce “Invisible Backdoor Attack against Self-supervised Learning” 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
Hanrong Zhang, Zhenting Wang, Boheng Li, Fulin Lin, Tingxu Han, Mingyu Jin, Chenlu Zhan, Mengnan Du and 2 more
A team developed INACTIVE, a backdoor attack for self-supervised learning models that implants visually imperceptible triggers, achieving average attack success rates of 99.09% while preserving model utility. The attack successfully evades current state-of-the-art backdoor defenses and common image corruptions across various SSL frameworks.
Key insights
- The primary reason existing imperceptible triggers fail in SSL is the 'entanglement' of backdoored sample distributions with the model's augmentation processes, which are designed to learn invariant features.
- By specifically designing triggers to operate in color spaces (HSV/HSL) and maximizing distributional distance from the SSL augmentation space, imperceptibility can be maintained while significantly enhancing attack effectiveness.
- A carefully balanced multi-objective loss function combining disentanglement, feature alignment, and comprehensive stealthiness metrics is crucial for creating robust, imperceptible, and effective backdoor attacks against SSL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.