Stealthy data poisoning attack method on offline reinforcement learning in unmanned systems

Aiming at the limitations in effectiveness and stealth of existing offline reinforcement learning(RL) data poisoning attacks, a critical time-step dynamic poisoning attack was proposed, perturbing important samples to achieve efficient and covert attacks. Temporal difference errors, identified through theoretical analysis as crucial for model learning, were used to guide poisoning target selection. A bi-objective optimization approach was introduced to minimize perturbation magnitude while maximizing the negative impact on performance. Experimental results show that with only a 1% poisoning rate, the method reduces agent performance by 84%, revealing the sensitivity and vulnerability of offline RL models in unmanned systems.