Calibrated Adversarial Sampling: A Bandit Approach to Multi-Attack Adversarial Training
A new paper on arXiv (2511.12265) introduces Calibrated Adversarial Sampling (CAS), a method to stabilize multi-attack adversarial training (AT) for deep neural networks (DNNs). The authors reformulate multi-attack AT as a multi-armed bandit optimization problem, sampling a single attack per iteration that balances exploration and exploitation. This reduces training cost and mitigates optimization conflicts and parameter drift. Experiments show CAS achieves superior overall robustness at low computational cost. The paper is a replace-cross announcement, indicating a revised version.
Key facts
- Paper arXiv:2511.12265v2, type replace-cross.
- Proposes Calibrated Adversarial Sampling (CAS) framework.
- Reformulates multi-attack adversarial training as a multi-armed bandit problem.
- Samples one attack per iteration, balancing exploration and exploitation.
- Reduces training cost and controls parameter drift.
- Achieves superior overall robustness in experiments.
- Addresses inefficiencies of computing all attacks and stochastic sampling.
- Published on arXiv, not yet peer-reviewed.
Entities
Institutions
- arXiv