Substra enables privacy-preserving federated learning for real-world healthcare data

AI Impact Summary

Substra is presented as an open-source federated learning framework designed for real-world production, enabling privacy-preserving ML by keeping data on local servers and only exchanging model weights. This approach reduces data exposure while potentially improving model robustness through diverse, multi-institution data sources. The MELLODDY collaboration demonstrates a concrete, large-scale use case where pharma competitors share insights without exposing raw data, with Hugging Face contributing to the ecosystem. While powerful, real-world deployment of federated learning remains technically complex, requiring careful integration with existing IT security and data governance controls.