Lightweight Federated Learning for Efficient Network Intrusion Detection

Network Intrusion Detection Systems (NIDS) play a crucial role in ensuring cybersecurity across various digital infrastructures. However, traditional NIDS face significant challenges, including high computational and storage costs, as well as privacy risks. To address these issues, we introduce a novel method called "Lightweight-Fed-NIDS," which harnesses federated learning and structured model pruning techniques for NIDS. The primary advantage of our contribution lies in the one-time computation of the pruning mask, without the need to access clients' data. This mask is then distributed to all clients and utilized to prune and optimize their local models. Furthermore, we leverage the power of Convolutional Neural Network (CNN) architectures, including ResNet-50, ResNet-101, and VGG-19, to extract essential features from raw traffic flows. We evaluate the performance of our method using various NIDS benchmark datasets, such as UNSW-NB15, USTC-TFC2016, and CIC-IDS-2017. Our technique achieves up to a 3X acceleration in training time compared to traditional, unpruned federated learning models, while maintaining a high detection rate of \sim ~99 %. Additionally, our method reduces model size by 90%, demonstrating its efficiency and scalability for real-world NIDS deployments. These results highlight the potential of Lightweight-Fed-NIDS to enhance network security while addressing privacy concerns and resource constraints in distributed environments.