Deep reinforcement learning-based automated stealthy wear attack on digital twins

Researchers propose a deep reinforcement learning (DRL)-based automated stealthy wear attack that covertly manipulates control signals in digital twin industrial systems to accelerate joint wear and evade anomaly detection. Benchmarks against TD3, SAC, PPO and A2C show that SAC is consistently best in sample efficiency, stability and attack effectiveness. In an evaluation in an industrial scenario with a UR10e robotic arm, the attack significantly raised the target joint torque, accelerating equipment degradation and driving up maintenance costs.