Causative attack to Incremental Support Vector Machine

The support vector machine (SVM), which is an effective and robust classifier, has been applied into many security problems successfully. Since these problems change over the time, the incremental learning is usually applied in SVMs. However, an intelligent adversary who misleads the learning of SVMs by manipulating the training samples may present in the security problems. As the conventional incremental learning algorithm of SVMs does not consider the adversarial attack, its performance may be affected significantly in the adversarial environment. In this paper, we investigate the vulnerabilities of incremental learning algorithm of SVM under the adversarial attack. The attack model to the incremental learning algorithm of SVM is proposed. The experimental results show that the accuracy of the incremental learning algorithm of SVM reduces dramatically under the proposed attack.