Perturbation-Controlled Deep Q-Learning With Human-Teaming for Enhancing Adversarial Robustness

Abstract

In this letter, an integrated framework of perturbation-controlled deep Q-network with human- teaming is proposed to effectively mitigate the impact of adversarial disturbances, specifically false data injection and denial-of-service attacks. Through a convergence and error-compensation mechanism, the proposed integration substantially reduces the effects of such errors. The incorporation of human intervention introduces a favorable trade-off between convergence speed and robustness, which is particularly critical in safety-sensitive applications where robustness must take precedence over fast convergence through adaptive quantized perturbation injection integrating with human-teaming. Consequently, the algorithm enables efficient and reliable recovery while maintaining satisfactory performance levels. Simulation results demonstrate that within adversarial intervals, the proposed method exhibits superior capability in mitigating and compensating for injected errors compared to conventional deep Q-network-based approach.