Federated Learning for Cybersecurity: A Privacy-Preserving Approach
Edi Marian Timofte, Mihai Dimian, Adrian Graur, Alin Dan Potorac, Doru Balan, Ionuț Croitoru, Daniel-Florin Hrițcan, Marcel Pușcașu
Abstract
The growing number of cyber threats and the implementation of stringent privacy regulations have revealed significant shortcomings in traditional centralized machine learning models, especially in distributed systems like the Internet of Things (IoT). This study presents a Federated Learning (FL) framework designed for intrusion detection and malware classification. This framework enables decentralized model training while preserving data locality and minimizing communication overhead. The proposed architecture incorporates lightweight, privacy-preserving techniques, including gradient clipping, differential privacy, and encrypted model aggregation, to ensure secure and efficient collaboration across heterogeneous clients. Experimental results on two widely adopted cybersecurity benchmarks demonstrate that the framework achieves detection accuracies above 90%, maintains privacy loss below 5%, and improves communication efficiency by over 25%. These results confirm the viability of FL as a scalable, privacy-compliant approach for next-generation cybersecurity systems in highly distributed infrastructures.