Edge-Native Machine Learning Frameworks For Real-Time Intrusion Detection in IoT Networks

Abstract

The rapid expansion of Internet of Things (IoT) devices has created vast networks across smart cities, industries, and healthcare. This expansion introduces severe security vulnerabilities. Traditional Intrusion Detection Systems (IDS) rely heavily on cloud computing. Cloud-based models face critical challenges, including high latency, bandwidth limitations, and privacy risks. To address these issues, edge-native machine learning offers an efficient alternative. This paper proposes a lightweight, edge-native IDS designed for real-time anomaly detection. We introduce a Lightweight Temporal Convolutional Network (L-TCN) optimized specifically for edge hardware. The proposed model uses dilated causal convolutions to capture complex temporal sequence patterns in network traffic without the immense parameter overhead of Transformers or deep LSTMs. We employ model pruning to reduce size and computation costs, ensuring deployment viability on constrained devices like the Raspberry Pi. The framework was evaluated across three benchmark datasets: NSL-KDD, CICIDS2017, and BoT-IoT. Experimental results show our model achieves up to 98.92% accuracy and an F1-score of 98.75%. Compared to cloud-based solutions, the edge-native framework reduces inference latency by 82% and cuts energy consumption by 45%. This study proves that optimized edge-native architectures provide superior real-time security while maintaining high classification fidelity.