A Novel Defense Mechanism Against Label-Flipping Attacks for Support Vector Machines

Datasets are important for machine learning for training, and they can be searched on the Internet these days. However, attackers may attack datasets by label flipping attacks. Therefore, a method to defend against the attack is needed. In this paper, we propose an algorithm with two processes, namely sanitizer and defender (SPD), to protect against uncertain datasets and label flipping attacks. We also conduct a series experiments to validate the relabeled dataset and maintain the accuracy of the support vector machine (SVM). Our experiment results show that the proposed SPD algorithm can effectively block the attack by relabeling the dataset and gain high accuracy with relabeled data.