Universal Adversarial Perturbations Against Machine Learning-Based Intrusion Detection Systems in Industrial Internet of Things

The security of the Industrial Internet of Things (IIoT) has emerged as a prominent concern in cyber-security due to the potential impact of attacks against IIoT on physical infrastructure. Machine learning-based intrusion detection systems recently have been demonstrated to be an effective tool for protecting the systems in IIoT. However, the vulnerability of machine learning-based intrusion detection systems to adversarial attacks hinders their further application in IIoT. This paper aims to further explore the adversarial attacks in IIoT to better evaluate the security of machine learning-based intrusion detection systems in this area. Our research primarily focuses on the generation of universal adversarial perturbations in IIoT, a topic that received limited attention in previous literature. Two novel attack methods based on a unified framework are proposed to utilize the original input-dependent adversarial perturbations of gradient-based or optimization-based adversarial attack methods to craft universal adversarial perturbations with better performance and transferability. The proposed attack methods conceal the underlying implementation details of the target attack methods, exploiting the original adversarial perturbations in a black-box manner. This enhances their flexibility, making them applicable in a wider range of scenarios and enabling them to be combined with most gradient-based or optimization-based attack methods. Comprehensive experiments are conducted on three mainstream intrusion detection datasets, i.e., NSL-KDD, Gas Pipeline, and Edge-IIoTset, to validate the effectiveness of the proposed methods. The preliminary experimental results demonstrate the feasibility of universal adversarial perturbations in IIoT and the superiority of the proposed methods to state-of-the-art attack methods.