Entropy-Guided Gradient Pruning with Informer for Intrusion Detection System in Internet of Things

Abstract

The internet of things (IoT) is becoming increasingly significant in computer networks and applications. The existing deep learning (DL)-based intrusion detection system (IDS) suffers from high computational complexity and poor generalization, where irrelevant network flows are due to the inability of the filter. To address this challenge, this research proposes a novel entropy-guided gradient pruning (EGGP) with an Informer for IDS classification. The EGGP eliminates the less impactful and redundant network flows based on entropy and gradients that focus on significant traffic. The EGGP is integrated with the Informer backbone, which utilizes ProbSparse self-attention and sequence distillation while effectively capturing long-term temporal dependencies. The experiments are conducted on the ToN-IoT, BoT-IoT, IoT-23, and CICIDS2019 datasets; the EGGP-Informer outperforms other state-of-the-art methods while maintaining lesser memory usage and inference time. Furthermore, this research includes preprocessing for data quality and class imbalance mitigation using class weights, thereby validating its robustness through an ablation study. Therefore, the EGGP-Informer achieves an accuracy of 99.98% for ToN-IoT, 99.99% for BoT-IoT, 99.96% for IoT-23, and 99.89% for CICIDS2019, demonstrating its scalability to diverse networks and uneven distribution of attacks.