Signal Asan Medical Center trained AI on 300,000 records across three hospitals without moving raw data

Summary

The report uses a Korean hospital study to show a practical use of the mathematics behind lattice-based cryptography. The same properties that resist quantum attacks enable fully homomorphic encryption, which allows calculations to run on data while it stays encrypted. In 2024, researchers at Asan Medical Center applied this method to train AI models on more than 300,000 patient records from three hospitals. At no point did any hospital's raw data leave its own servers. The resulting models performed better than anything a single institution could have built alone. The case suggests that sensitive data can be pooled for analysis without being exposed, a capability that classical cryptography cannot match.