A Regularization Method for Enhancing the Robustness of Regression Networks

Neural networks are widely used in control systems to learn dynamics and controllers from the sampled data. However, they are susceptible to adversarial attacks; a small perturbation in the input may cause a significant change in the output, posing serious security risks. Most of the existing work focuses on the robustness of neural network classification tasks. Since learning dynamics and controllers are regression tasks, ensuring the robustness of regression networks is a significant and practically valuable problem. In this work, we explore adversarial attacks against regression networks and propose a defense mechanism based on regularization. We introduce a penalty designed to promote modest changes in the output of the learning function as the input undergoes variations. To evaluate the performance of our regularization method, experiments were carried out on regression datasets and control systems. The simulation results show that our proposed regularization term can effectively improve the robustness of the regression network.