Privacy-Preserving Federated Deep Learning for Cybersecurity Analytics in Decentralised IoT Networks

Abstract

Rapid propagation of Internet of Things (IoT) devices has created new opportunities and threats, particularly when they are operated in decentralized networks where data aggregation to a central point is difficult or privacy- sensitive. In this work, we investigate a privacy-preserving FDL framework for cybersecurity analytics in decentralised IoT systems. In contrast to centralised data collection which is used by classical machine learning models, the proposed scheme allows the training of deep neural networks in a distributed manner among IoT nodes without disclosing raw data and thereby preserving user and device privacy. The study combines federated averaging, differential privacy and homomorphic encryption in order to reduce adversarial threats, for better protection against inference attacks while maintaining high accuracy for detecting anomalies and cyber intrusions. Simulation results over diverse IoT setups show robust convergence of the proposed approach with low communication overhead, and outperforming standard centralised and the non-federated model in terms of accuracy, resiliency, and privacy guarantees. The results emphasise the potential of federated deep learning as a linchpin for secure and scalable trustworthy cybersecurity analytics in future decentralised IoT networks.