D2.1 Privacy Preserving Feature Selection, Classificationand Federated Learning

The HARPOCRATES project aims to develop solutions for private and secure data analysis and sharing. This deliverable (D2.1 Privacy Preserving Feature Selection, Classification and Federated Learning) presents the outcomes of research conducted in Work Package 2 (WP2). WP2 has focused on advancing privacy-preserving machine learning (PPML) techniques to address one of today’s most pressing challenges in data science: how to learn from data without exposing it. In an era where AI systems rely on large volumes of personal and often sensitive data, safeguarding privacy during machine learning tasks is critical—particularly in domains such as healthcare and cyber threat intelligence. This deliverable covers three key areas: feature selection, classification, and federated learning—each enhanced with novel approaches to improve efficiency, scalability, and data protection. Feature Selection: We introduce a lightweight, secure protocol using multi-party computation. The method significantly reduces the complexity of mutual information estimation, achieving more than 1,000× speed-up in regression tasks while preserving model performance. This makes privacy-preserving data preparation practical and efficient. Classification: In distributed environments, split learning allows a client and server to train a model jointly without sharing raw data. However, it still risks leaking sensitive intermediate information. We designed two secure split learning protocols based on homomorphic encryption and function secret sharing, ensuring that intermediate outputs remain protected—even from untrusted servers. Federated Learning: We improved the state of the art in privacy-preserving federated setups. Our work optimised existing homomorphic encryption schemes such as Flashe and introduced a new version, Flashev2, to better support encrypted training. We also addressed the often-overlooked problem of secure hyperparameter tuning in federated learning by proposing PRIVTUNA, a framework that enables collaborative tuning using multiparty homomorphic encryption. This approach provides high accuracy and efficiency without privacy loss. Additionally, we explored the use of differentially private synthetic data for time series anomaly detection. Our results show that incorporating differential privacy into the data generation process can preserve strong utility, in some cases outperforming non-private synthetic data.