Action-Manipulation Attacks Against Stochastic Bandits: Attacks and Defense [article]

Guanlin Liu, Lifeng lai
2020 arXiv   pre-print
Due to the broad range of applications of stochastic multi-armed bandit model, understanding the effects of adversarial attacks and designing bandit algorithms robust to attacks are essential for the safe applications of this model. In this paper, we introduce a new class of attack named action-manipulation attack. In this attack, an adversary can change the action signal selected by the user. We show that without knowledge of mean rewards of arms, our proposed attack can manipulate Upper
more » ... ence Bound (UCB) algorithm, a widely used bandit algorithm, into pulling a target arm very frequently by spending only logarithmic cost. To defend against this class of attacks, we introduce a novel algorithm that is robust to action-manipulation attacks when an upper bound for the total attack cost is given. We prove that our algorithm has a pseudo-regret upper bounded by O(max{log T,A}), where T is the total number of rounds and A is the upper bound of the total attack cost.
arXiv:2002.08000v2 fatcat:ywgcxdifwnck3mpxxjsln5zsk4