ZKML
Zero-Knowledge Machine Learning. A cryptographic technique that lets a prover convince a verifier that a specific model produced a specific output, without revealing the model weights or the input.
Also known as: zk-ML, zero-knowledge ML, verifiable inference
ZKML applies zero-knowledge proof systems to machine-learning inference. The setup has three roles: a prover that runs the model on some input and produces a proof, a verifier that checks the proof, and a public commitment to the model itself. If the proof verifies, the verifier learns that the prover ran the specific committed model on the input and got the specific output. The verifier learns nothing about the weights, nothing about the input beyond what the output reveals, and does not need to re-run the model. That is the property that makes ZKML interesting for on-chain AI: a smart contract can trust an inference result without trusting the operator that produced it.
The economics are still rough. Producing a ZK proof for a forward pass through a small neural network can take minutes; for a multi-billion-parameter language model, hours or days. The verifier-side cost is small (often a few milliseconds), which is what makes the on-chain verifier possible, but the prover-side cost is the bottleneck on every production deployment. Three project families are pushing on this in different directions. Halo2-based SNARKs (EZKL) trade trusted-setup for proving speed. Custom STARKs (Modulus Labs, Giza’s LuminAIR with Starkware’s S-two) trade proof size for transparency and no trusted setup. opML (Ora Protocol) sidesteps the proof cost entirely by using optimistic verification, where proofs only run when a result is disputed.
In DeAI, ZKML is the cleanest answer to a hard question: how do you trust a decentralised inference network without trusting the operators? The TEE answer (Phala, Nillion, Intel SGX) gives you hardware-rooted attestation; the ZKML answer gives you cryptographic attestation. TEEs are practical today at full model scale; ZKML is practical today only for small models or simple operations. The ZKML projects are running multi-year roadmaps to close the gap. Giza’s LuminAIR phases its way through primitive operators, then fused operators, then an on-chain Cairo verifier. The shipping product almost always uses TEEs or optimistic verification today, with ZKML positioned as the longer-term endgame.