arXiv
AI/ML
When your AI model gets sick from bad teammates — and how to vaccinate it
Imagine a classroom where half the students deliberately give you wrong answers. A new technique lets one teacher at the front spot the liars and correct the class average in real time, even when the honest students' homework looks nothing alike.
This means federated learning (where phones and hospitals train AI without sharing raw data) can now survive when over half its participants are compromised — a threshold experts thought impossible.
Bug reported: No