# Vulnerability of Deep Reinforcement Learning to Policy Induction Attacks

## TL;DR

This paper demonstrates that Deep Q-Networks in reinforcement learning are vulnerable to adversarial input attacks, which can manipulate policies, and introduces a novel transferability-based attack method validated through experiments.

## Contribution

It reveals the vulnerability of DQNs to adversarial attacks and proposes a new transferability-based policy induction attack method.

## Key findings

- Adversarial examples transfer across different DQNs.
- Transferability enables effective policy manipulation.
- Experimental results confirm attack efficacy in game scenarios.

## Abstract

Deep learning classifiers are known to be inherently vulnerable to manipulation by intentionally perturbed inputs, named adversarial examples. In this work, we establish that reinforcement learning techniques based on Deep Q-Networks (DQNs) are also vulnerable to adversarial input perturbations, and verify the transferability of adversarial examples across different DQN models. Furthermore, we present a novel class of attacks based on this vulnerability that enable policy manipulation and induction in the learning process of DQNs. We propose an attack mechanism that exploits the transferability of adversarial examples to implement policy induction attacks on DQNs, and demonstrate its efficacy and impact through experimental study of a game-learning scenario.