Lightweight Deep Learning Models for Real-Time Edge Intrusion Detection

The increasing use of edge devices in critical domains such as healthcare, industry, and smart infrastructure is making them an attractive target for cyber attackers. To prevent cyberattacks, Intrusion Detection Systems (IDS) based on Machine Learning (ML) have been widely studied but they often face challenges in scalability, adaptability, and generalization to complex attack patterns. Deep Learning (DL) models solve many of these issues with higher accuracy, but they are difficult to deploy in resource-constrained edge devices. Thus, this work aims to investigate optimized DL models that can be deployed on edge devices. We benchmark four representative architectures: DNN, CNN, LSTM, and GRU, using the CICIoT2023 dataset across binary, 8-class, and 34-class classification scenarios. We evaluate accuracy, precision, recall, F1-score, inference time, CPU/RAM usage, and model size for each of them. Our results show that pruning and quantization compress models by a factor of 4 to 10 and reduce inference latency by up to 14 times while maintaining accuracy above 98% for binary classification and giving competitive performance in multi-class settings. These findings demonstrate that optimized DL models can enable real-time, energy-efficient IDS on edge devices without compromising effectiveness. We argue for a shift from conventional DL toward optimized DL for edge security, ensuring scalable and practical protection in resource-constrained environments.