Zero-knowledge machine learning (zkML)

A machine learning model is ultimately a long sequence of arithmetic operations. zkML tools translate those operations into a circuit or a program for a zero-knowledge virtual machine, run the model and generate a succinct proof that the computation was carried out correctly. Anyone, including a smart contract, can verify that proof quickly, while chosen parts such as the weights or the user's data stay hidden.

The appeal for crypto is trust without disclosure. A protocol could accept a credit score, a fraud check or a trading signal on-chain knowing it came from an agreed model, an identity system could confirm a biometric match without exposing the image, and AI agents could prove they followed a stated policy. It is one of several routes to verifiable inference, alongside secure hardware and re-execution by multiple parties.

Proving is still very expensive: generating a proof can take orders of magnitude longer than running the model normally, so practical uses focus on small models or parts of larger ones. A proof shows that a specific model ran correctly on given inputs, not that the model is accurate, fair or fed truthful data.