Hebert Silva^1,2 and Regina Moraes^1,3, ¹Universidade Estadual de Campinas, Brazil, ²National Industrial Training Service, Brazil, ³University of Coimbra, Portugal

The CSAI-4-CPS model leverages federated learning to collaboratively train machine learning models, providing accurate and up-to-date results while preserving data privacy. This approach is particularly beneficial in complex and dynamic Cyber-Physical Systems (CPS) environments where traditional centralized machine learning models may fall short. This paper presents the first validation of the CSAI-4-CPS model using a framework implemented for an IoT system and describes the new features of its expanded version. Real-time threat detection, consideration of false positives, and verification and validation of results on nodes that benefit from federated learning are among these new capabilities. It also compares the results obtained with and without the model. IoT systems often represent the most challenging scenarios in CPS cybersecurity, and in most cases, IoT devices are part of a more complex CPS structure, where they are usually the most vulnerable assets. The application of CSAI-4-CPS to predict malicious traffic in Internet of Things (IoT) networks appears promising. The results demonstrate that the model effectively detects intrusions in these datasets. By employing federated learning and a self-adaptive architecture, the model maintains its accuracy and relevance as new data emerges.

Federation Learning, Data Privacy, Cybersecurity, CPS, IoT