One of the biggest challenges in
#AI and
#healthcare is how to collaboratively train large AI models from multiple hospitals without compromising patient privacy?
Thrilled to introduce our new paper, DeCaPH, which is now officially published in
@TheLancet @eBioMedicine!
**Decentralised, Collaborative, and Privacy-preserving Machine Learning for Multi-Hospital Data**
What does DeCaPH do?
🔒 Privacy-preserving collaborative ML: Train together without sharing sensitive data.
🌐 Decentralization: No central server is needed, reducing privacy risks.
🚫 Robust against privacy attacks: Designed for healthcare's unique privacy needs.
**Highlights**
--- Tailored for healthcare setting: Designed specifically for hospitals, DeCaPH caters to their unique threat models and privacy requirements, achieving the optimal utility-privacy trade-off.
--- Potential Real-world Impact: We conducted comprehensive evaluations on three cross-silo datasets for three distinct tasks: clinical outcome prediction from electronic health records, cell classification with single-cell RNA transcriptomics, and pathology identification using chest radiology images.
--- Robust to privacy attacks: Our evaluations show that models trained with DeCaPH are more resilient against privacy attacks than those without privacy protection.
Kudos to Emmy Fang (
@EmmyFangCY), who is an amazing PhD student cosupervised by me and
@NicolasPapernot , for her incredible leadership in this project!
Read More:
Paper:
sciencedirect.com/science/ar…
arXiv:
arxiv.org/abs/2402.00205
Code:
github.com/cleverhans-lab/De…
@VectorInst @UofTCompSci @UofT_LMP @UHN @bradwouters @drbarryrubin