Adaptive Lightweight Intrusion Detection for IoT using Partial–Temporal Convolution and Knowledge Distillation

Abstract

A security tool can be used to monitor network related tasks and activities of a system to identify unwanted activities and unauthorized access that can be referred to as intrusion Detection System (IDS). Within the Internet of Things (IoT) systems, the IDS examines the large volume and non-uniform traffic of the distributed and resource-constrained devices with the aim of detecting the cyber-attacks in real time. However, the existing designs of the IoT-based IDS are not only expensive to compute but are also limited in their dynamism to the variations in the traffic, and degrade substantially in their performance under the condition of the distribution shifts. In an attempt to relieve these problems this paper will propose a new light and adaptive IDS architecture. Adaptive Feature-aware Traffic Encoding (AFTE) is the method which dynamically adjusts the network parameter, which is the feature of the network the network according to the statistic and time, which enhances resistance to the changing IoT traffic. It has been demonstrated that Hybrid Partial -Temporal Convolution Network (HPTCN) provides a good way to capture discriminative spatio-temporal features at a low cost of computation through partial convolution and gated temporal modelling. Continual Multi-source Knowledge Distillation (CMKD) is an online cross-domain adaptation algorithm that employs knowledge transfer between two or more teachers that adapt with time, and thus is more resistant to unseen and few-shot attacks. It is experimentally demonstrated that the proposed framework is very precise in finding and much less expensive in computation and enhanced generalization in dynamic IoT.