Prior-enhanced Semi-supervised Federated Learning for IoT Intrusion Detection: A Game Theory and Comparative Learning-Based Approach

Federated Learning (FL) has gained prominence as a method for intrusion detection in Internet of Things (ID-IoT) edge devices, with a central server, aiming to address privacy concerns. However, FL-based intrusion detection faces challenges, including privacy risks from transmitting model parameters and difficulties in handling non-Independently Distributed (non-IID) data. Recent studies propose semi-supervised FL approaches using knowledge distillation to address these issues. However, such methods require the central server to distribute a common public dataset to each client, leading to slower cold starts due to variations in client speeds. To address these challenges, we propose Prior-enhanced Semi-supervised Federated Learning for ID-IoT. Specifically, we leverage comparative learning and K-means clustering to generate prior pseudo-labels for public datasets, expediting the cold-start process. Experimental evaluation on real-world datasets with three non-IID scenarios demonstrates the effectiveness of our approach.