Backdoor Attack and Defense Methods for AI–Based IoT Intrusion Detection System

Abstract

The Internet of Things (IoT) is an emerging technology that has attracted significant attention and triggered a technical revolution in recent years. Numerous IoT devices are directly connected to the physical world, such as security cameras and medical equipment, making IoT security a critical issue. Artificial intelligence (AI) based intrusion detection technology for IoT can rapidly detect network attacks and improve security performance. However, this technology is vulnerable to backdoor attacks. As an important form of adversarial machine learning (ML), backdoor attacks can allow malicious traffic to evade detection of the intrusion detection system, posing a significant threat to the IoT security. This study focuses on backdoor attack and defense methods for AI–based IoT intrusion detection system. Specifically, we first use different ML and deep learning (DL) classification models to classify IoT traffic data, thereby achieving intrusion detection within IoT. Additionally, we employ data poisoning techniques to implant backdoors into models, enabling backdoor attacks on classification models. For backdoor defense, we propose backdoor detection and mitigate methods: (1) The proposed backdoor detection method is achieved by leveraging the strong correlation between the backdoor trigger and the target classification; (2) we utilize the unlearning method to mitigate the backdoor effect, enhancing the robustness of classification networks. Extensive experiments were conducted on the CICIOT2023 dataset to evaluate the effectiveness of IoT intrusion detection, backdoor attack, and defense.