Secure Federated Learning for Multi-Party Network Monitoring

Network monitoring is essential for IT infrastructure health, enabling proactive threat detection, bandwidth optimization, and data analysis. Federated Learning facilitates collaboration and expands threat detection capabilities by allowing multiple clients to train models while preserving privacy and security without compromising sensitive information. This work explores the application of Federated Learning for secure network monitoring by investigating its use in various data partitioning settings to train Deep Learning models across multiple data partitions, incorporating secure aggregation to enhance data privacy. We propose an approach for multiparty training of deep neural networks for time series-based network load forecasting with secure aggregation. Our approach shows that a collaboratively trained model, on horizontal partitions, performs 11%-14% better than model training only on a single partition of that type. The model trained on vertical partitions achieved performance comparable to that of the models trained on a complete data set. Finally, examining the proposed approach on horizontal and vertical partitions proved its viability in both aggregation settings, regular and secure. These contributions demonstrate the feasibility of Federated Learning to improve interoperability between multiple organizations while addressing privacy and security concerns.