STEM与日常科技·英语30篇（5）

17 / 30

联邦学习如何让医院数据不出本地却提升AI模型？

Federated learning trains AI models across hospitals without moving patient records from secure on-site servers.
Each hospital runs computations locally, then shares only encrypted model updates—not raw images or notes—with a central server.
The server aggregates these updates using weighted averaging, reinforcing patterns common across diverse populations.
Since data never leaves local firewalls, HIPAA and GDPR compliance stays intact without sacrificing model accuracy.
Differences in scanner types or diagnosis protocols are naturally absorbed during aggregation, improving generalizability.
Doctors see faster tumor detection in X-rays because models learn from Boston, Berlin, and Bangalore—without cross-border transfers.
Bandwidth needs drop over ninety percent compared to uploading full datasets to cloud clusters.
Adversarial checks verify update authenticity, blocking corrupted or malicious contributions before model merging.
Real-world pilots show pneumonia classifiers improve sensitivity by seven percentage points after three federation rounds.
This approach turns data silos into collaborative intelligence—without centralization or compromise.