SLAB: Stealthy Deep Reinforcement Learning Backdoor Attack with Latent Space Trigger

Deep reinforcement learning, as a crucial deep learning technology, has demonstrated significant performance in numerous real-world domains in recent years. However, research indicates that deep reinforcement learning is vulnerable to backdoor attacks. As a stealthy attack method, backdoor attacks are only activated under specific conditions, severely compromising the security of the model. Existing backdoor attack methods on single-agent DRL generally focus on embedding backdoor triggers by directly applying perturbations in the pixel space, guiding the model to learn backdoor policies through modifying the data features of original states. However, such pixel-space perturbation methods tend to cause significant differences in visual features between backdoor states and normal states, resulting in poor concealment. To address this limitation, this paper proposes a backdoor attack method that modifies latent space vectors as backdoor trigger. Compared with existing methods, our approach does not directly modify the pixel space of the environment, thus making it more stealthy and natural. Experiments conducted using the DQN and PPO algorithms in Atari environments demonstrate the effectiveness of our method.