Policy Poisoning in Batch Reinforcement Learning and Control

TL;DR

This paper investigates security vulnerabilities in batch reinforcement learning and control, demonstrating how small data modifications can manipulate learned policies, with a unified attack framework applicable to various models.

Contribution

It introduces a unified convex optimization framework for policy poisoning attacks in batch RL and control, applicable to multiple standard models, with analysis of attack feasibility and cost.

Findings

Policy poisoning can effectively manipulate learned policies.

The attack framework is convex and guarantees global optimality.

Experiments confirm the effectiveness of the proposed attacks.

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.