Lightweight model-contrastive federated learning with multi-center clustering for IoT intrusion detection

With the rapid development of the Internet of Things (IoT), the data privacy and security issues arising from its widespread application are becoming increasingly prominent. Traditional centralized intrusion detection systems (IDSs) face the challenges of data privacy leakage and high communication overhead. Federated learning (FL), as a distributed machine learning paradigm, effectively protects data privacy through the mechanism of “data doesn't move, model moves”, but still suffers from the problems of data heterogeneity, inefficient communication, and insufficient privacy protection in practical applications. To address the above shortcomings, this paper proposes a lightweight multicenter clustering model comparison federation learning framework (MCFL-MC) to trade-off robustness and privacy in non-independently identically distributed (non-IID) environments. MCFL-MC achieves adaptive modeling of data distribution and dynamic migration of clients through a two-phase dynamic clustering algorithm based on the spectral clustering gap statistics and mitigates the problem of data heterogeneity at the same time. On this basis, we introduce a multi-source temporal model comparison learning method and add historical global models as negative samples, which significantly improves the model convergence speed and generalization performance. To reduce the communication overhead, we design an adaptive gradient compression strategy based on cosine annealing mechanism. In addition, MCFL-MC combines both differential privacy and homomorphic encryption techniques to construct a secure and reliable multicenter aggregation mechanism. The experimental results show that MCFL-MC achieves 97.29 % and 99.64 % accuracy on ToN-IoT and Edge-IIoT datasets, respectively, which outperforms the existing federated learning methods, and excels in communication efficiency and privacy protection.