Near ZKML

Recently, with progress in AI, large language models get often integrated in crypto applications. A possible use case would through oracles. They, however, can represent a week point in the otherwise trustworty crypto infrastructure. Therefore, proof of Inference - i.e a confirmation that relevant computations actually took place could be used to prevent fraud. That is why we decided to use EZKL library in our ZKML project. The main idea is to create an oracle that can give a guarantee described above. That guarantee is a proof given by the EZKL library, in which specific inputs, outputs and the parameters of the model can (but don't have to be) taken into account. Later they are just being verified and returned back by the oracle. For example, one can ensure that a particular GPT model gives a particular output. The inputs could be sensitive so do necessarily have to be taken in the proof. This has a wide range of potential applications, for example on protocols concentrating on AI as Near.

To build this project, we integrated several technologies to create a seamless workflow. We used Ezkl for zero-knowledge proof generation, ensuring the integrity and privacy of model inputs, outputs and parameters. We used Python for backend part and JavaScript and React for frontend. The smart contract is written in Rust and could easily be uploaded on the Near blockchain. In our project a variety of different models is supported (i.e the full process of proof generation and verification is possible).