ATS-O2A: A state-based adversarial attack strategy on deep reinforcement learning

•An effective and stealthy adversarial attack method on deep reinforcement learning.•A new attack effect measurement index for attacked effectiveness and stealthiness.•Experiment tests address the proposed method is better than the other two. In recent years, deep reinforcement learning has been widely applied in many decision-making tasks requiring high safety and security due to its excellent performance. However, if an adversary attacks when the agent making critical decisions, it is bound to bring disastrous consequences because humans cannot detect it. Therefore, it is necessary to study adversarial attacks against deep reinforcement learning to help researchers design highly robust and secure algorithms and systems. In this paper, we proposed an attack method based on Attack Time Selection (ATS) function and Optimal Attack Action (O2A) strategy, named ATS-O2A. We select the critical attack moment through the ATS function, and then combine the state-based strategy with the O2A strategy to select the optimal attack action which has profound influence as targeted action, finally we launch an attack by making targeted adversarial examples. In order to measure the stealthiness and effectiveness of the attack, we designed a new measurement index. Experiments show that our method can effectively reduce unnecessary attacks and improve the efficiency of attacks.