In recent years, Artificial Intelligence (AI) has seen a remarkable surge in adoption in many everyday applications, primarily fueled by Machine Learning (ML) algorithms that rely on extensive data for model training. However, privacy constraints and the decentralization of data across various repositories, often constrained by data sharing limitations, present a significant challenge. To address this issue, Federated Learning (FL) techniques have emerged with the promise of facilitating collaborative model training across disparate devices or entities while offering better data privacy guarantees. While it is true that FL enhances data privacy, security concerns still remain, including privacy attacks that compromise the confidentiality of training data like attribute inference and even data reconstruction attacks. To strengthen the baseline privacy provided by FL, it is essential to research and develop novel privacy enhancement methods for Federated Learning. Our goal is to deliver a highly scalable, armored Federated AI service platform for researchers, enabling AI-powered studies of multi-site, siloed, cross-domain, cross-border European datasets with high privacy guarantees which comply with data privacy regulations such as the General Data Protection Regulation (GDPR). This paper explores how FL can be a useful tool for implementing collaborative training of ML models with privacy guarantees, focusing on the main challenges to be addressed to achieve an armored FL framework.
Ortega-Fernandez et al. (Thu,) studied this question.