[LUM#22] Towards better protected data

While data is the raw material of artificial intelligence, it can also be particularly sensitive, especially in the healthcare sector. How can we reconcile AI, confidentiality and sovereignty? For Aurélien Bellet, researcher at the Institut Desbrest d'épidémiologie et de santé publique1the solution lies in federated learning.

In order to include a large number of patients in clinical studies and make them more meaningful, research uses multicenter studies. In other words, they involve several hospitals or clinics at the same time, sometimes even in several different countries. Advantage: this method enables large-scale studies to be carried out on patients from a wide range of social and geographical backgrounds.

Sharing without sharing

This organization also has a drawback: multicenter studies require health data from several institutions to be pooled on a single server, "which makes it difficult to maintain control over the data, and could also jeopardize their confidentiality", explains Aurélien Bellet, a researcher at Idesp. How can medical research implement these collaborations while reducing the risk of sensitive information being divulged? One solution is to share... without sharing. This is federated learning. " This means that each institution's data can be processed on site, without having to exchange, assign or transmit it," explains the federated learning specialist.

To meet this challenge, the researchers are creating learning algorithms capable of operating on data stored locally, rather than centralized on a server, as is the case with conventional machine learning methods. "It is then the intermediate results of this learning that are exchanged as and when required, rather than the data itself. In this way, we alternate between local learning and aggregation of results", explains Aurélien Bellet , who is working with the university hospitals of Lille, Caen, Amiens and Rouen.

Democratization

To promote confidentiality and respect for medical ethics, federated learning "is part of the solution, even if it is often not sufficient to guarantee data confidentiality", explains the researcher whose team is also working with the Commission Internationale Informatique et Liberté (CNIL ) on the complex issue of data protection.

Because the benefits of federated learning go far beyond the medical field. "It can also be of interest to companies wishing to collaborate with competitors without giving them access to certain sensitive information," adds Aurélien Bellet, who also sees federated learning as an opportunity to make AI more accessible. "It's a form of democratization of artificial intelligence and machine learning, because it doesn't require investment in a large infrastructure, opening the way to collaborative uses, for example by citizen collectives."

Also watch :

Aurélien Bellet's conference on federated learning

UM podcasts are now available on your favorite platforms (Spotify, Deezer, Apple podcasts, Amazon Music...).

  1. Idesp (Inserm, UM)