Policy Poisoning in Batch Reinforcement Learning and Control

Abstract

We study a security threat to batch reinforcement learning and control where the attacker aims to poison the learned policy. The victim is a reinforcement learner / controller which first estimates the dynamics and the rewards from a batch data set, and then solves for the optimal policy with respect to the estimates. The attacker can modify the data set slightly before learning happens, and wants to force the learner into learning a target policy chosen by the attacker. We present a unified framework for solving batch policy poisoning attacks, and instantiate the attack on two standard victims: tabular certainty equivalence learner in reinforcement learning and linear quadratic regulator in control. We show that both instantiation result in a convex optimization problem on which global optimality is guaranteed, and provide analysis on attack feasibility and attack cost. Experiments show the effectiveness of policy poisoning attacks.

NeurIPS 2019

Cited by in corpus (10)

• Robust Reinforcement Learning on State Observations with Learned Optimal Adversary
• Poisoning Attack against Estimating from Pairwise Comparisons
• Provably Efficient Black-Box Action Poisoning Attacks Against Reinforcement Learning
• Policy Teaching in Reinforcement Learning via Environment Poisoning Attacks
• Reward Poisoning in Reinforcement Learning: Attacks Against Unknown Learners in Unknown Environments
• Disturbing Reinforcement Learning Agents with Corrupted Rewards
• Multi-Task Federated Reinforcement Learning with Adversaries
• Fair Clustering Using Antidote Data
• Environment Shaping in Reinforcement Learning using State Abstraction
• Adaptive Teaching of Temporal Logic Formulas to Learners with Preferences