Stealthy Adversarial Attacks on Stochastic Multi-Armed Bandits

TL;DR

This paper investigates stealthy reward poisoning attacks on stochastic multi-armed bandit algorithms, revealing conditions under which such attacks can almost always succeed and highlighting security vulnerabilities.

Contribution

The study introduces a new notion of stealthy attacks on MABs, analyzes their success conditions, and demonstrates their potential to bypass existing detection methods.

Findings

Stealthy attacks depend on environmental conditions and initial rewards.

Most existing attacks are detectable due to aggressive reward manipulations.

Stealthy attacks can almost always succeed against certain MAB algorithms.

Abstract

Adversarial attacks against stochastic multi-armed bandit (MAB) algorithms have been extensively studied in the literature. In this work, we focus on reward poisoning attacks and find most existing attacks can be easily detected by our proposed detection method based on the test of homogeneity, due to their aggressive nature in reward manipulations. This motivates us to study the notion of stealthy attack against stochastic MABs and investigate the resulting attackability. Our analysis shows that against two popularly employed MAB algorithms, UCB1 and $\epsilon$-greedy, the success of a stealthy attack depends on the environmental conditions and the realized reward of the arm pulled in the first round. We also analyze the situation for general MAB algorithms equipped with our attack detection method and find that it is possible to have a stealthy attack that almost always succeeds. This brings new insights into the security risks of MAB algorithms.