Federated learning theoretical tools and platform development

High-Performance Computing (HPC) systems play a crucial role in modern scientific and industrial applications, yet their growing complexity introduces challenges in fault detection and system reliability. This thesis explores the application of Federated Learning (FL) and Federated Transfer Learning (FTL) to enhance anomaly detection in HPC nodes. Traditional machine learning (ML) approaches require large labeled datasets, which are often unavailable in HPC environments. To address this, we propose decentralized learning methodologies that leverage data across multiple nodes while preserving data privacy and reducing network overhead. Our research investigates the effectiveness of unsupervised, semi-supervised, and supervised FL-based anomaly detection models, including LSTM autoencoders and other deep learning techniques. Empirical evaluations on real-world supercomputing systems, such as Marconi100 at CINECA, demonstrate substantial improvements in anomaly detection performance, with F1-score enhancements from 0.31 to 0.84 in semi-supervised learning and a 15-fold reduction in data collection time. Fur-thermore, FTL enables knowledge transfer from high-performing to low-performing nodes, reducing the need for extensive retraining. By integrating FL and FTL, this thesis ontributes to the advancement of intelligent, scalable, and efficient anomaly detection in HPC systems, facilitating improved system uptime, performance, and reliability.