PPFL-DCS: Privacy-Preserving Federated Learning Using Neural Transformer and Leveraging Dynamic Client Selection to Accommodate Data Diversity

The vulnerabilities and security issues of industrial Cyber-Physical Systems (CPSs), such as Intrusion Detection Systems (IDSs), have significantly increased due to the rapid integration of conventional industrial setups with advanced networking and computing technologies like 5G, software-defined networking, and artificial intelligence. Coping strategies for such challenges frequently involve transferring data to a central location, which raises concerns about latency, efficiency, and privacy. To address these issues, Federated Learning (FL) was developed as a solution to mitigate both the privacy concerns of organizations and the complexities of networked systems. However, FL-based techniques still have shortcomings, FedAvg equally weights weak models, risking suboptimal results; FL also faces Membership Inference privacy attacks. To address these challenges, we propose PPFL-DCS, an FL framework that incorporates a weighted mechanism for dynamic client selection, accounting for the performance of each local model and data size of each client in integration with a Neural Transformer System (NTS) that enhances the system‘s robustness against the MIA attacks. The NTS limits the impact and gains of attackers, thereby reducing the effectiveness of MIAs. Extensive experiments demonstrate that PPFL-DCS achieves a high detection accuracy of 97.424% for cyber threats in industrial CPSs, and highlight its efficiency over state-of-the-art techniques.