Enhancing Adversarial Robustness for SVM Based on Lagrange Duality

Adversarial examples generated by adversarial attacks bring security threats to the application of machine learning models. Certified defense can improve the adversarial robustness of the model against various adversarial attacks. Despite substantial research efforts to enhance the adversarial robustness of models in recent years, the focus has mainly been on deep neural networks. However, it is crucial to extend this research to include classic models such as support vector machines (SVM), which remain important even in the era of deep learning. Therefore, the issue of certified defense of SVM needs to be paid attention to. In this paper, a verified training SVM method (VT-SVM) based on Lagrange duality is proposed. The proposed method incorporates adversarial robustness into the SVM learning framework. Experimental results are provided to demonstrate the efficacy of the proposed method in ensuring both accurate predictions and enhanced adversarial robustness of the model.