secure pdm: a novel byzantine fault tolerant federated learning framework using a robust pca-based anomaly detection approach

With the proliferation of federated learning programs as a suitable framework for protecting user privacy and reducing the computational overhead of ai algorithms, various industries have also turned to the widespread use of this framework in industrial applications such as improving predictive maintenance (pdm). however, despite its increasing applications, several security challenges, such as byzantine attacks, make the application of federated learning in industries questionable. byzantine attacks in fl can degrade model performance by injecting malicious updates, causing model divergence or biased learning. this reduces accuracy, and can introduce security vulnerabilities such as backdoors. to address this problem, we propose a byzantine fault tolerant (bft) federated learning algorithm designed to improve pdm in industrial applications. our proposed approach uses a pca-based anomaly detection algorithm to detect and mitigate local byzantine updates. also, a game theory-based reward mechanism is designed to promote honest participation and discourage malicious behavior among federated users. the proposed framework is evaluated using the predictive maintenance datasets ai4i 2020 and nasa acoustics and vibration. the results show that our proposed framework effectively detects and mitigates byzantine attacks, enhancing the overall reliability of pdm in industrial applications.