Reproduce “Universal Adversarial Attack on Attention and the Resulting Dataset DAmageNet” 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
Sizhe Chen, Zhengbao He, Chengjin Sun, Jie Yang, Xiaolin Huang
Researchers at Shanghai Jiao Tong University developed "Attack on Attention (AoA)", a method that crafts universal adversarial examples by manipulating a deep neural network's attention map, yielding significantly higher transferability across diverse models. They also introduced DAmageNet, a dataset of 50,000 such examples from ImageNet, demonstrating its effectiveness against various undefended and adversarially trained models.
Key insights
- DNNs, despite varied architectures, exhibit common vulnerabilities in their attention mechanisms, which can be exploited to craft highly transferable adversarial examples.
- Manipulating a network's internal attention map more effectively enhances the transferability of adversarial examples compared to directly attacking only the classification output.
- Combining attention-based loss functions with existing transferability-enhancement techniques (like Scale Invariance) leads to state-of-the-art black-box attack performance against diverse models and defense mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.