A²HD: Adaptive Adversarial Training for Hyperdimensional Computing-Based Intrusion Detection Against Adversarial Attacks

The Internet of Things (IoT) systems possess various security vulnerabilities due to increased inter-connectivity. Machine learning-based intrusion detection system (ML-IDS) is an IoT security approach to monitor system data and detect malicious activity. However, deploying ML solutions in IoT environments is challenging due to their substantial computational requirements. Hyperdimensional computing (HD) is an efficient learning approach and can be an alternative ML-IDS solution. Nevertheless, its prediction performance can be impacted by adversarial attacks. In this paper, we propose an adaptive adversarial training defense framework, A<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>HD, to protect HD-based intrusion detection against adversarial attacks. A<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>HD discovers the most effective adversarial attack and incorporates the attack samples into our adaptive adversarial training process. Our results show the effectiveness of the proposed defense, where we improve HD prediction performance by up to 145% and 35% with respect to no defense and the state-of-the-art adversarial training defense under various adversarial attacks.