Researcher Echo Wang from CrossSBlockingce explains what ZKML is, why it should be used in ML, and looks forward to the future of ZKML. Despite facing challenges such as “parameter distortion” and “high computing power requirements for large models,” ZKML is expected to unlock the key to the cross-integration of “Web3 + AI.”
What is ZKML? ZK (Zero-Knowledge) is a concept in cryptography that refers to a type of proof or interactive process. The prover can prove the truthfulness of a statement to the verifier without revealing any specific information about the statement. ML (Machine Learning) is a subfield of AI. Machine learning learns from input data, summarizes it, and forms models to make predictions and decisions. Most machine learning models today are neural networks, so we mainly refer to machine learning as neural network machine learning. (Note: When we talk about ZKML, we are talking about zero-knowledge proofs for the inference step of creating ML models, not the training of ML models)
Why use ZK in ML? 1) The potential applications of zero-knowledge cryptography can help us determine whether specific content in the model inference of machine learning is generated by applying a specific model to a given input. 2) Zero-knowledge circuits can provide a way to verify whether the outputs of machine learning models such as GPT-4, text-to-image models like DALL-E 2, or any other model represent what they were created for.
We can summarize the above as a machine learning trust framework, in which multiple levels of machine learning must be trusted for the entire machine learning process to be trusted: (1) Model: the type and algorithm of the model must use the correct computational model. Model parameters should be transparent or democratically generated in some cases, but are not easily falsifiable in all cases. (2) Integrity: the input data has not been tampered with or contaminated by malicious input, and has undergone appropriate preprocessing. The model runs correctly, can produce normal results, and can correctly create, maintain, and manage parameters. (3) Privacy: input data will not be actively disclosed. The model itself or the computation process will not be made public. The model owner keeps the parameters confidential.
What is the future of ZKML? ZKML combines the privacy protection and verification capabilities of zero-knowledge proofs with the data processing and decision-making capabilities of machine learning: ZK helps ML solve the trust proof problem and provides an on-chain environment; mature AI technology ML helps ZK achieve Web3 ecology expansion and application innovation. Although facing challenges such as “parameter distortion” and “high computing power requirements for large models”, it is still expected to unlock the key to the cross-fusion of “Web3+AI”.