Learning-Based Defense Against Prediction-Based Displacement Attacks on Formation Control

Formation control is a widely performed task in multi-agent systems. However, agents under formation control are susceptible to attacks, and how to defend against the attack is important for system security. This paper aims to design a new method to protect the system from prediction-based displacement attack, which predicts the movement of agents, diverts them from their original tracks and drives them to a preset trap. The defense is challenging due to the robustness of the attacker's prediction mechanism. Also, undermining the attack can be costly and harm the system's performance. To address this issue, we first build obstacle-avoidance and formation control models of the holonomic and non-holonomic agents and specify the prediction-based attacks to limit the cases where the defense strategy can be effective. Then, we propose a learning-based defense strategy, which learns the prediction mechanism of the attackers. It modifies the control input based on self-adaptive fuzzy methods to reduce prediction accuracy and guarantee that the agents can reach the goal. Simulation results demonstrate that our method is better than the previous noise-based one at reducing the accuracy of the attacker's prediction.