LOD-DLIDS: A Lightweight Optimization-Driven Deep Learning Framework for Intrusion Detection in IoT Networks

Abstract

The proliferation of Internet of Things (IoT) networks has exacerbated security threats, especially in scenarios with limited computational resources, where traditional intrusion detection systems cannot be deployed. Deep learning (DL) methods have powerful detection performance but are often too large for real-world deployment on edge devices. In this paper, we propose a Lightweight Optimization-Driven Deep Learning Intrusion Detection System (LOD-DLIDS) for real-time IoT security. The proposed framework employs a lightweight convolutional neural network (CNN) to extract local features and a lightweight Vision Transformer to consider global traffic information. A cross-attention fusion is used to improve feature discrimination between diverse attack types. To facilitate deployment, structured pruning and quantization-aware training are adopted to achieve a compact model size and low computation cost without compromising performance. Experiments are conducted on a series of attacks from the BoT-IoT dataset, such as denial-of-service, reconnaissance, information-theft and keylogging. Evaluation results show that LOD-DLIDS achieves 98.56% accuracy and a 98.28% F1-score with low latency and footprint. This confirms the model provides a good trade-off between accuracy and efficiency, and is applicable to resource-constrained IoT systems.