Exploring Hyperdimensional Computing for Anomaly-Based Intrusion Detection Systems

AUGUSTO CESAR DA F. DOS SANTOS, Leandro Santiago de Araujo, Claudio Miceli de Farias, Leopoldo André Dutra Lusquino Filho

Keywords: Hyperdimensional Computing, Intrusion Detection System, IoT security, Anomaly Detection, Machine Learning, NSL-KDD, UNSW-NB15, CIC-IDS-2017, BCCC-CIC-IDS2017, BotNetIoT

Abstract: Traditional Intrusion Detection Systems (IDS) are often inefficient at protecting Internet of Things (IoT) devices due to the computational constraints of these devices and the evolving nature of cyberattacks. Although classical Machine Learning (ML) algorithms such as Random Forest and Decision Tree achieve high accuracy, their computational cost and memory consumption make them impractical for large-scale deployment in resource-constrained environments. To tackle this issue, this research investigates the use of Hyperdimensional Computing (HDC), a bio-inspired and energy-efficient alternative, as the core of an IDS for IoT. We propose a comprehensive benchmarking analysis comparing six state-of-the-art HDC models (BinHD, NeuralHD, OnlineHD, AdaptHD, DistHD, and CompHD) against six traditional ML algorithms on modern cybersecurity datasets (NSL-KDD, IoT-Flock, UNSW-NB15, BCCC-CIC-IDS2017, CIC-IDS-2017, and BotNetIoT-L01). Our results demonstrate that the most advanced techniques in HDC models are able to improve accuracy while maintaining low memory consumption, especially NeuralHD and AdaptHD, which obtained the best accuracy results among the HDC models. NeuralHD consumed 467x less memory with a 4.72% accuracy difference compared to Random Forest on the UNSW dataset. We conclude that HDC emerges as a viable and necessary paradigm for IoT IDS, providing an ideal trade-off between performance, computational cost, and memory consumption.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 9817