Reproduce “Understanding Model Ensemble in Transferable Adversarial 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
Wei Yao, Zeliang Zhang, Huayi Tang, Yong Liu
This research presents a theoretical framework explaining the effectiveness of model ensemble adversarial attacks in generating transferable adversarial examples. It decomposes transferability error into vulnerability and diversity, and derives an upper bound for this error, supporting practical guidelines that increasing the number and diversity of surrogate models, and reducing their complexity, enhances attack transferability.
Key insights
- Transferability error can be mathematically decomposed into vulnerability and diversity, offering a new analytical perspective on the factors influencing transferable adversarial attacks.
- The derived theoretical upper bound indicates that increasing the number of surrogate models, enhancing their diversity, and judiciously reducing their complexity can improve adversarial transferability.
- The mathematical constructs and derived guidelines demonstrate a strong analogy between adversarial transferability and model generalization, validating a long-standing observation in the field.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.