Multi-party homomorphic encryption for precision medicine

Researchers from EPFL and CHUV have developed an encryption technique capable of advancing precision medicine. Also involving teams from MIT and Harvard, this research led to the development of a solution to the main obstacle to improving diagnostics and targeted therapies: the sharing of sensitive data.

Current limited and risky approaches

To progress, large-scale biomedical research must bring together clinical data from hospitals, practices and clinics around the world, and use it to create federated machine learning models. Hence a risk of data leakage. However, according to the EPFL press release, some of the current approaches only offer limited protection of patient privacy in this context, by forcing institutions to share intermediate results. While other approaches sacrifice the precision of the results, introducing noise into the data to limit potential leaks. To address these challenges, the researchers developed the FAMHE system, a multi-party homomorphic encryption approach applied to federated learning. According to EPFL, the FAMHE system is a good compromise between data protection, the accuracy of research results and practical computing time.

Calculations on data without communication between the parties

In their article, published in the journal Nature Communications, the researchers explain that their approach combines the power of homomorphic (HE) encryption “to perform calculations on encrypted data without communication between the parties, with the advantages of interactive protocols that can simplify considerably some expensive HE operations. ” The FAMHE system introduces a new approach where each participating institution performs local calculations and encrypts intermediate results using multi-party homomorphic encryption. The results are then combined and distributed to each institution for further calculations. “By sharing only encrypted information, our approach ensures that, when necessary, a minimum level of obfuscation can be applied only to the end result to protect it from inference attacks, rather than being applied to all results. intermediaries ”, specify the researchers in their article.

As effective as less secure approaches

The effectiveness of the FAMHE system was demonstrated by correctly reproducing the results of the two studies which relied on the prior transfer and centralization of data. “So far, no one has been able to replicate studies showing that federated analysis works at scale. Our results are precise and were obtained with a reasonable calculation time ”, explains Jean-Pierre Hubaux, professor at EPFL and lead author of the study. Researchers are already in advanced discussions, with partners on a global scale, to deploy FAMHE on a large scale. “We want this to be integrated into the normal activities of medical research”, concludes Jean Louis Raisaro, doctor of the CHUV and one of the principal researchers of the study.

Listed in the last Gartner Emerging Technology Hype Cycle, homomorphic encryption is being explored by tech giants. But also by la start-up Inpher, partly based at EPFL, whose solutions also combine secure multi-party computing, federated learning and fully homomorphic encryption.

> On the subject: Confidential IT is revealed

See also  'Nintendo will release an improved Switch console in 2021' | NOW

Leave a Comment