A Hybrid Autoencoder Transformer Federated Learning Model for Secure IoT Intrusion Detection

Abstract

The explosive growth of IoT networks has also raised the threat of distributed cyber attacks, and efficient privacy-preserving intrusion detection is even more important. The traditional centralized IDS models are faced with some performance limitations in terms of data privacy and non-scalable solutions that could work under heterogeneous devices. To cope with these issues, this paper suggests a hybrid Autoencoder– Transformer federated learning system for the secure IoT intrusion detection. The Autoencoder is responsible for latent features extraction, meanwhile the Transformer captures spatio-temporal traffic patterns and trained jointly in a privacy-preserving federated learning framework with FedProx optimization and differential privacy. The ToN-IoT-IDS dataset is used in our approach evaluation. Experimental results demonstrate the effectiveness of our model by achieving 97.23% accuracy, 96.5% macro-F1 and 97.3 ROC-AUC for CECIC task, compared to state-of-the-art solutions. The experimental results show that our approach has better detection performance, stable convergence, and enhanced privacy-preserving for distributed IoT scenarios.