A Privacy-Preserving Hybrid Intrusion Detection System for IoT Networks Using Federated and Deep Learning Models

Abstract

The exponential increase in IoT devices has truly brought in security challenges never noticed before from heterogeneous traffic patterns with limited resources and f]rom ever-changing cybercriminal behaviors. This study proposes a privacy-preserving Intrusion Detection System (IDS) for IoT environments through the integration of FL with sophisticated DL models. The framework utilizes AE, DNN, and a hybrid AE+CNN architecture for anomaly detection and intrusion classification. Model evaluation was conducted using N-BaIoT, a dataset comprising traffic from different IoT devices under genuine attack scenarios. Data preprocessing steps such as normalization, encoding, and feature engineering were carried out to improve data quality and reduce noise. The FL paradigm allows distributed training on IoT devices without any exposure of raw data to any intermediary, thereby strengthening privacy and scaling. Experimental results show that AE obtains an accuracy of 95% and strong anomaly-detection abilities, whereas FL+DNN attains an accuracy of 90.39% and higher precision (97.99%) for classifying known attacks. The hybrid AE+CNN bettered all other models with 96.5% accuracy maintained a balance on recall and precision and showed great robustness toward zero-day and complex attacks. Comparative analysis reveals that the more light-weight nature of AE+CNN makes it suitable for edge and defense against imbalance more, while FL+DNN is better suited to privacy-aware distributed settings. This study exemplifies how unsupervised and supervised DL techniques under FL offer a powerful potential to develop efficient, scalable, and accurate IDS solutions. Future work will try to direct false positives reduction, explainable AI, and feasibility enhancement to deployments on real-world IoT systems.