Abstract
Reinforcement learning (RL) for the Markov Decision Process (MDP) has emerged in many security-related applications, such as autonomous driving, financial decisions, and drone/robot algorithms. In order to improve the robustness/defense of RL systems against adversaries, studying various adversarial attacks on RL systems is very important. Most previous work considered deterministic adversarial attack strategies in MDP, which the recipient (victim) agent can defeat by reversing the deterministic attacks. In this paper, we propose a provably ``invincible'' or ``uncounterable'' type of adversarial attack on RL. The attackers apply a rate-distortion information-theoretic approach to randomly change agents' observations of the transition kernel (or other properties) so that the agent gains zero or very limited information about the ground-truth kernel (or other properties) during the training. We derive an information-theoretic lower bound on the recipient agent's reward regret and show the impact of rate-distortion attacks on state-of-the-art model-based and model-free algorithms. We also extend this notion of an information-theoretic approach to other types of adversarial attack, such as state observation attacks.
Abstract (translated)
针对马尔可夫决策过程(MDP)中的强化学习(RL),在许多安全相关应用中已经取得了广泛应用,例如自动驾驶、金融决策和无人机/机器人算法。为了提高面对对手时的鲁棒性和防御能力,研究各种对RL系统的对抗攻击是非常重要的。以往大多数工作都考虑了确定性的对抗性攻击策略,在这种情况下,接收者(受害者)可以通过逆转这些确定性攻击来击败它们。在这篇论文中,我们提出了一种理论上“无敌”或“不可计数”的类型对抗攻击方法。 攻击方采用速率失真信息理论的方法随机改变代理对转换核(或其他属性)的观察结果,使得代理在训练期间几乎无法获得有关真实内核(或其他属性)的有效信息。我们推导了接收者代理人回报遗憾的信息论下限,并展示了这些率失真攻击对最先进的基于模型和不基于模型算法的影响。此外,我们还扩展了这种信息理论方法到其他类型的对抗性攻击,例如状态观察攻击。 这种方法为防御强化学习系统面临的复杂安全挑战提供了一种新的视角,强调了在设计鲁棒的AI和机器学习解决方案时考虑随机、非确定性的干扰的重要性。
URL
https://arxiv.org/abs/2510.13792