Reproduce “Scaling Laws for Black box Adversarial Attacks” 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
Chuan Liu, Huanran Chen, Yichi Zhang, Jun Zhu, Yinpeng Dong
Researchers from Tsinghua University empirically discovered a log-linear scaling law for black-box adversarial attacks, demonstrating that attack success rate increases linearly with the logarithm of surrogate model ensemble size. This methodology achieved high attack success rates on state-of-the-art MLLMs, including 85.0% on GPT-4o, and revealed a robustness hierarchy among commercial models.
Key insights
- A universal log-linear scaling law governs black-box adversarial attacks, where the Attack Success Rate (ASR) increases linearly with the logarithm of the ensemble cardinality (T).
- Resolving the 'gradient conflict' inherent in ensemble methods is critical for unlocking and observing the benefits of scaling the number of surrogate models.
- As ensemble size increases, adversarial perturbations evolve from unstructured noise to semantically meaningful features of the target class, indicating the distillation of robust, common weaknesses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.