Lightweight Federated Few-Shot Learning-Based Network Intrusion Detection for Resource-Constrained IoT Devices

Resource-constrained IoT devices are particularly vulnerable to network intrusions, highlighting the need for effective security mechanisms. Network Intrusion Detection Systems (NIDS) address these threats by monitoring traffic and detecting malicious behavior. However, deploying anomaly-based NIDS on such devices faces three key challenges: privacy concerns in centralized training, the scarcity of labeled data and high computational demands of machine learning models in real-world IoT environments. To address these challenges, this study presents a Lightweight Federated Few-Shot Learning (Light-FFSL) framework, specifically designed for secure and efficient intrusion detection in IoT settings. The framework preserves data privacy by employing federated learning, which enables collaborative model training across devices without exposing sensitive data. To overcome data scarcity, an LSTM-based prototypical few-shot model is used, allowing the system to generalize from only a small number of labeled samples per class. To reduce computational cost, the framework incorporates dynamic scheduling and convergence-based updates. These mechanisms dynamically select participating devices and reduce unnecessary communication for federated training. In addition, the model architecture integrates pruning and post-training quantization into an LSTM-driven prototypical network. Pruning removes redundant parameters before training to reduce complexity, while quantization converts the trained model into an efficient integer-based form, lowering memory requirements and speeding up inference. The effectiveness of the proposed framework is validated through experiments on real-world IoT network datasets, demonstrating its ability to operate efficiently under resource constraints.