Poisoning machine learning based wireless IDSs via stealing learning model

Abstract

Recently, machine learning-based wireless intrusion detection systems (IDSs) have been demonstrated to have high detection accuracy in malicious traffic detection. However, many researchers argue that a variety of attacks are significantly challenging the security of machine learning techniques themselves. In this paper, we study two different types of security threats which can effectively degrade the performance of machine learning based wireless IDSs. First, we propose an Adaptive SMOTE (A-SMOTE) algorithm which can adaptively generate new training data points based on few existing ones with labels. Then, we introduce a stealing model attack by training a substitute model using deep neural networks (DNNs) based on the augmented training data in order to imitate the machine learning model embedded in targeted systems. After that, we present a novel poisoning strategy to attack against the substitute machine learning model, resulting in a set of adversarial samples that can be used to degrade the performance of targeted systems. Experiments on three real data sets collected from wired and wireless networks have demonstrated that the proposed stealing model and poisoning attacks can effectively degrade the performance of IDSs using different machine learning algorithms.