A Federated Learning Scheme for Eavesdropper Detection in B5G - IIoT Network Orientations

In heterogeneous Beyond 5G (B5G) environments, the adoption of sophisticated Physical Layer (PHY) enhancements, such as Physical Layer Security (PLS) techniques, play a vital role in safeguarding sensitive transmissions against eavesdropping threats. However, traditional security mechanisms face scalability and privacy limitations. Machine and Deep Learning (ML/DL) have been proposed to address these challenges, because of their potential to assist in real-time security-aware decision making at the PHY. However, training such models in a centralized manner is simply not feasible or scalable due to the huge amount of data generated by the participating devices as well as due to privacy concerns. These limitations become more pronounced in B5G-Industrial Internet of Things (IIoT) settings, where highly dynamic topologies, massive numbers of resource-constrained devices, strong interference, and non-IID data distributions pose unique challenges to centralized security schemes, such as increased latency, poor generalization and vulnerability to data privacy breaches. In this context, this work introduces a Federated Learning (FL)-based framework designed to detect active eavesdroppers in the uplink of B5G IIoT networks using Channel State Information (CSI). A simulation framework where malicious users are strategically placed based on transmission characteristics is proposed, and CSI is collected to generate a realistic dataset. To this end, three DL models for eavesdroppers detection are evaluated: two Deep Convolutional Neural Networks (DCNNs) with different input features and a Long Short-Term Memory (LSTM) network. All models are trained and tested within both Centralized Learning (CL) and FL settings, for various numbers of clients, aiming to maximize the overall system’s secrecy throughput. According to the presented results, the proposed approach achieved an average detection accuracy over 90% for the best-performing model and an average secrecy rate of only 1.5 bps below the corresponding CL model. Moreover, the training time was reduced by approximately 20%. These results establish a lightweight, and highly effective solution for securing B5G IIoT networks.