Federated Learning with Differential Privacy for Intrusion Detection in Internet of Flying Things: A Robust Approach

In recent times, Federated Learning (FL) has emerged as a decentralized framework for intelligent knowledge sharing, demonstrating a degree of privacy preservation in safeguarding users’ sensitive information across various cyber-physical networks like the Internet of Flying Things (IoFT) network. Nevertheless, there exists a vulnerability wherein adversaries can deduce clients’ gradient or parameter updates to compromise privacy. This vulnerability is particularly concerning due to the clients’ utilization of cybersecurity models aimed at securing the network against cyber intrusions and attacks. This study investigates the utilization of Gaussian and Laplace differential privacy (DP) mechanisms by clients during local training to obfuscate their model parameters, thereby mitigating the risk of data leakage. Extensive simulations utilizing the edge-IIoT dataset, validate the effectiveness of perturbing client models, in terms of upholding privacy and enhancing global model performance. Thus, demonstrating significant global model accuracy of 90%, specifically with the introduction of Laplace noise outperforming an unperturbed global model in a scalable network.