Review for NeurIPS paper: Adversarial Attacks on Linear Contextual Bandits

Summary and Contributions: Summary & Contributions: Authors study the scopes of adversarial attacks in linear contextual bandit algorithms which have applications in a wide range of domains. The authors consider adversarial attacks on both reward and the context and analyze the robustness (or lack of it) of various contextual linear bandit algorithms including LinUCB, LinTS, epsilon-greedy etc. Empirical evaluations are presented on various synthetic and two real datasets to examine the effect of attacks on these algorithms. Strengths: The problem of analyzing the effect of adversarial attacks on bandit algorithms are indeed interesting and well motivated, and present work is supposedly the first one to analyze this for stochastic contextual linear bandits. Authors also analyze some popularly studied bandit algorithms, like LinUCB, LinTS, epsilon-greedy, and showed the attacking strategies (as optimization problems) to fool above algorithms for playing some targeted suboptimal arm majority number of times. Experiments are fairly detailed and reported on a large set of datasets showing the effect on learning rate of existing techniques on different degree+type of attacks.