Apprentissage automatique fédéré pour l’IA collaborative dans le secteur de la santé

Federated learning (FL) represents today a key working paradigm to empower AI research while guarantying data governance and confidentiality through decentralized learning applications. FL enables different clients to jointly learn a global model without sharing their respective data, and is therefore particularly suited in AI applications with sensitive data, such as in healthcare. Nevertheless, the use of FL in the medical domain is currently at its infancy, with only a handful of pioneering applications demonstrated in real world conditions. One of the critical aspects about the application of FL in real world conditions concerns the aspects of security and safety. Ill intentioned parties may interfere during the federated learning procedure to degrade/modify models performances, or retrieve information about other clients? data. There is currently a greyzone of potential privacy threats associated to the development and exploitation of complex AI methods to sensitive data. These threats arise anytime we can interfere with the model training or exploitation processes, to gather more information about the data used to generate such a model. In this work, we provide an overview of current research and challenges on the security and safety of federated learning, with special focus on healthcare application.