Reproduce “Yet Another Intermediate-Level 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
Qizhang Li, Yiwen Guo, Hao Chen
Researchers from ByteDance AI Lab and UC Davis introduced a method that enhances the transferability of adversarial examples in black-box scenarios by leveraging the entire optimization process of baseline attacks and a learned linear mapping between intermediate features and adversarial loss. The approach consistently improves attack success rates against various unseen models compared to existing transfer-based methods.
Key insights
- A linear relationship exists between intermediate-level feature discrepancies and the adversarial loss, which can be leveraged for improving transferability.
- Utilizing information from earlier iterations of a baseline attack's optimization process can lead to better transferability than using only the final result.
- Intermediate-level normalization significantly enhances the stability and performance of the proposed method.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.