Fed-Deep CTRL (Light GBM)-Scalable, Interpretable, and Privacy-Preserving Federated Learning Framework for Wireless Intrusion Detection System

Abstract

A learning-based intrusion detection system automates the understanding and reporting of network traffic information. The systems using neural networks and machine learning have demonstrated good detection accuracy, but the accuracy is entirely reliant on the type and volume of data. Additionally, there are issues with its scalability, privacy, efficiency, and interpretability. With the innovative fed-deep CTRL (light GBM) method, this study focuses on creating a local-global federated architecture. On the federated learning framework, a novel method was developed using a mix of deep CTRL and Light GBM. Multiple clients in fed-deep CTRL handle the extraction of local attack data features that the server uses to train the detection to become more effective. The fed-deep CTRL was tested using the AWID dataset and its results were validated individually for each attack type. The overall accuracy for Amok was 97.21%, ARP was 96.42%, the beacon was 98.31%, caffe latte was 97.81%, CTS was 100%, death was 98.92%, disasso was 97.99% and evil twin was 98.46%. The overall accuracy of the model was observed to be 99.60%, the training overhead with 4 clients was 3.28 sec and the approach was shown to be highly interpretable, scalable, and secure.