Privacy-Prioritized Model Aggregation in ICPS: A Novel Approach to Federated Learning Aggregation With Lime and Blockchain

This paper contributes to the fields of Federated Learning (FL) and Industrial Cyber-Physical Systems (ICPS) privacy. It introduces a novel model aggregation technique aimed at prioritizing privacy protection for sensor data collected by Integrated Sensing Digital Devices (ISDD) during the aggregation process. By incorporating Lime, a local explanation technique, and Blockchain technology, the approach enhances both transparency and security in the global model update process. Furthermore, the implementation of transfer learning strengthens the adaptability of attack detection systems to evolving threats within the dynamic ICPS landscape. The paper also proposes a comprehensive privacy evaluation method, providing a systematic assessment of privacy measures within the FL context. Comparative evaluations against FedAVG underscore the superior adaptability, accuracy, and privacy enhancement capabilities of the proposed Lime AGG model, particularly in scenarios involving previously unseen attacks which is evaluated by CICIDS 2017 and 2018 datasets.