Integrating Machine Learning and Deep Learning for Robust and Privacy-Preserving Intrusion Detection: A Unified Framework and Experimental Study

An Intrusion Detection System (IDS) is still considered a necessary tool for ensuring the security of modern, highly distributed environments. However, the conventional methods have difficulty achieving all of the following: high detection accuracy, being robust against adversarial attacks, and ensuring privacy of data at the same time. Here, a synergistic hybrid solution that combines deep feature extraction via Convolutional Neural Networks (CNN) and Bidirectional Long Short Term Memory (BiLSTM) with ensemble based classification using Random Forest is proposed by the authors. Moreover, the solution uses federated learning to achieve decentralized training without sharing the local raw network traffic among the parties. The experimental results on the CICIDS2017 dataset show statistically significant performance improvements compared to deep learning baseline models $(p<0.01$) while obtaining a detection accuracy of 99.95% and very low variance over multiple runs. The authors have also subjected their model to FGSM, PGD, and C&W adversarial attacks and have observed that their model is more resilient than the single model counterparts. In addition, the cross dataset validation performed on UNSW NB15 confirms the generalization ability. The study results suggest that hybridizing representation learning, ensemble classification, and privacy preserving techniques yield a robust, practical, and stable intrusion detection system suitable for the present day network environments.