Learning to Attack: A Bandit Approach to Adversarial Context Poisoning

Researchers developed AdvBandit, a new black-box adversarial attack method that can exploit neural contextual bandits by poisoning context data without requiring access to internal model parameters. The attack uses bandit theory and inverse reinforcement learning to adaptively learn victim policies and optimize perturbations, achieving higher victim regret than existing methods.

• →AdvBandit introduces a novel black-box attack against neural contextual bandits that requires no internal model access.
• →The attack formulates context poisoning as a continuous-armed bandit problem with theoretical guarantees.
• →A surrogate model is constructed using maximum-entropy inverse reinforcement learning from observed context-action pairs.
• →Experiments on real-world datasets demonstrate superior attack performance compared to state-of-the-art baselines.
• →The research includes attack-budget control mechanisms to limit detection risk and computational overhead.