Robust intrusion detection based on personalized federated learning for IoT environment

In the dynamic and complex realm of the Internet of Things (IoT) and artificial intelligence (AI), it is a significant challenge to design a network intrusion detection system that balances accuracy, efficiency, and data privacy. Federated learning offers a solution by enabling the sharing of high-quality attack samples to enhance local models’ intrusion detection capabilities without compromising local data privacy. However, most existing research on federated learning for intrusion detection assumes homogeneity among local models, which can reduce detection accuracy in real-world scenarios where local datasets are often non-independent and identically distributed (Non-IID). The Non-IID characteristic, marked by varied distributional properties and correlations, impacts model convergence and stability. To address this challenge, we propose a personalized federated cross learning framework (pFedCross) for intrusion detection, to manage imbalanced and heterogeneous data distributions. First, we present a collaborative model cross aggregation algorithm for personalized local model update, to solve the problem that one global model cannot always accommodate all the incompatible convergence directions of local models. Then, we introduce a gradient approximation $\alpha$-fairness algorithm for global model generation to achieve a well-generalization. Finally, the experiments show that pFedCross outperforms baseline methods in improving model accuracy and reducing loss, highlighting its promise for enhancing IoT security.