Svetlana Boudko

To ensure data consistency and control, data centralization is a preferred solution for training machine learning models. However, data protection regulations, e.g. GDPR, as well as industrial competition, impose restrictions on information sharing among different organizations and individuals. Furthermore, this approach is technically challenging since the cost of collecting, storing, and processing all data in one centralized location is often prohibitively high. Google proposed federated learning for the collaborative training of machine learning models, aiming to handle the exchange of privacy-sensitive information in distributed environments and to reduce data transmission costs. In contrast to traditional machine learning, federated learning does not require local data to be collected, stored, and processed on a central server. Instead, this method enables on-device model training using client-specific data, with the obtained local model updates further aggregated on a central server. However, federated learning is not without its own privacy concerns, including risks of data leakage and inference attacks. To address these challenges, research is being conducted into various strategies, such as homomorphic encryption. By combining federated learning and homomorphic encryption, we can train machine learning models on encrypted data from different sources, thereby ensuring better data protection. The model never sees the raw data, only the encrypted version, and yet it can still learn from it. However, homomorphic encryption is computationally intensive and can significantly slow down the training process. In this talk, I take a look at the issues and prospects arising from the intersection of federated learning and multi-key homomorphic encryption, two advanced techniques in the field of secure and collaborative machine learning.

4th Workshop on Cybersecurity in Healthcare 4.0