Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection
This addresses vulnerabilities in widely used stochastic bandit algorithms for real-world systems, making it domain-specific but not incremental as it introduces a new practical model.
The paper tackled the problem of adversarial attacks on stochastic bandits by proposing a practical threat model called Fake Data Injection, which allows attackers to inject limited fake feedback samples, and showed that this can mislead algorithms like UCB and Thompson Sampling into selecting a target arm in nearly all rounds with sublinear attack cost.
Adversarial attacks on stochastic bandits have traditionally relied on some unrealistic assumptions, such as per-round reward manipulation and unbounded perturbations, limiting their relevance to real-world systems. We propose a more practical threat model, Fake Data Injection, which reflects realistic adversarial constraints: the attacker can inject only a limited number of bounded fake feedback samples into the learner's history, simulating legitimate interactions. We design efficient attack strategies under this model, explicitly addressing both magnitude constraints (on reward values) and temporal constraints (on when and how often data can be injected). Our theoretical analysis shows that these attacks can mislead both Upper Confidence Bound (UCB) and Thompson Sampling algorithms into selecting a target arm in nearly all rounds while incurring only sublinear attack cost. Experiments on synthetic and real-world datasets validate the effectiveness of our strategies, revealing significant vulnerabilities in widely used stochastic bandit algorithms under practical adversarial scenarios.