Application of Federated Learning in IOT Security for Signature Based Attacks

With the rapid growth of Internet of Things devices, a significant amount of security and privacy challenges arises, particularly regarding sensitive personal information. Traditional approaches of centralized machine learning are subject to this area of risk. While Federated Learning helps to alleviate some of those concerns through decentralized training of models, it is still vulnerable to model updates that may be used as an inference attack. This paper contributes a framework for privacy-preserving Federated Learning and discusses the use of Differential Privacy to provide data protection and reliability in heterogeneous IoT environments. The various models like Long Short-Term Memory, Gated Recurrent Unit, Tab Transformer, and Neural Network architectures are trained on different clients, and the resulting model is aggregated on central server using different federated learning aggregation strategies. Differential Privacy using Gaussian noise is used to communicate the weights to the central sever thereby increasing the privacy of the data transmitted. The proposed work is evaluated using the NSS-KDD dataset to detect Denial of Service, Probe, Remote to Local, User to Root based on their signatures. The experimental evaluation shows that the LSTM model achieves the highest global accuracy, making it more effective for privacy-preserving intrusion detection while also proposing its viability as a practical and usable reference model in real-world IoT security deployments in applications such as smart homes, industrial systems or critical infrastructure.