A covert attack detection strategy combining physical dynamics and effective features-based stacked transformer for the networked robot systems

Networked robots are vulnerable to malicious covert attack, which malevolently manipulates sensor and controller data without authorization and compromises the security of the robot’s physical process seriously. In response, this paper proposes a novel intrusion detection strategy using physical dynamics and an effective features-based stacked Transformer network (PD-EFST-IDS), to reveal the stealth of incursion with the assistance of unfalsifiable current data. The proposed IDS takes the predicted errors of the EFST network as the benchmark, and reports abnormal activity when the actual errors between predictions of the PD and the measurements deviate from the reference patterns. Specifically, the simplified dynamics model is initially established based on the estimated parameters screened by the orthogonal trigonometric decomposition, and the errors between the predicted torques and the credible measurements can be obtained. Subsequently, to solve the problem of PD model error and the tiny attack payload further being submerged due to uncertain factors such as disturbances and unmodeled factors, an EFST network is proposed for reconstructing torque error in joint space. Wherein the feature reconstruction module is prepared for embedding the critical current feature missing under attack, making it possible to maintain prediction accuracy even under malicious attacks. Then, we invoke another Transformer module for predicting errors, and it is further compared with actual ones to identify abnormal trajectories. The feasibility of EFST was demonstrated in actual experiments. The detection in three attack scenarios, and quantitative experiments examine our detection strategy.