zkML，人工智能之后的下一个宏大叙事？

Original Author: hitesh.eth

Original Translation: Frank, Foresight News

zkML, perhaps the next grand narrative after artificial intelligence.

However, zkML can be a bit complex for many people to understand, so in this article I will explain it in the simplest way possible.

What is zkML?

In short, zkML = ZKP + ML

where: ZKP = Zero Knowledge Proof, ML = Machine Learning.

Therefore: zkML = Zero Knowledge Proof Machine Learning

In a nutshell, it is using ZKP technology on machine learning models to generate output without revealing the sensitive data used in the training process, while ensuring the accuracy of the computation.

So what is a machine learning model? A machine learning model is a computer program that can make predictions based on a large amount of data.

For example, large language models like ChatGPT are built on top of machine learning models.

And what is inference? Inference is the process of analyzing user prompts, understanding the context, and using trained data models to provide results.

Let's take ChatGPT as an example:

The first step in the inference process is to write an input, such as the prompt "Write a Drake-style encrypted rap song".

The second step, ChatGPT will analyze the context, "Drake-style encrypted rap song". Then, based on the user's prompt, it will activate the training model, identify patterns in the training data, and create a Drake-style encrypted rap song as the output.

What can zkML do?

Throughout the entire inference process, there are two privacy issues that involve the potential leakage of sensitive data:

• Membership Inference attacks: Attackers can analyze the model's output to infer whether specific data points are part of the training process;

• Model Inversion attacks: Attackers may attempt to reconstruct fragments of training data from the output by constructing specific prompts;

What can zkML provide in this regard? zkML allows for inference on sensitive data without exposing the training data itself.

This is achieved by using ZK proof systems such as Plonky and Halo 2, with Plonky 2 currently being the fastest ZK proof system.

With zkML, attackers will never have direct access to the training data.

The current development status of zkML

As of now, zkML is still in its early stages, and several startups are working on building zkML infrastructure.

Risc Zero, in collaboration with Spice AI, is developing a complete zkML solution for developers.

Ingonyama is developing hardware specifically for ZK technology, which may lower the entry barrier into the ZK technology field and zkML could also be used in the model training process.

Modulus is using zkML to apply artificial intelligence to on-chain inference processes, and they currently have six partners who are building different zkML use cases:

For example, Upshot has built a price prediction model, Worldcoin is using Modulus for privacy identity verification, and AI ARENA uses zkML in the economic models of games.

Privacy-focused blockchain projects such as Oasis Protocol, Secret Network, and Aleo are also exploring use cases based on zkML in their ecosystems, and NOYA.ai is using zkML to build full-chain DeFi strategies.

OraProtocol is building a ZK-based trustless machine learning inference protocol where developers will be able to use zkML inference to build any decentralized application driven by machine learning and secured by Ethereum.

The entire narrative around zkML is still in its early stages, but I anticipate a hype cycle for this narrative to emerge in the coming months, given the bullish market. Therefore, now is an excellent time to closely follow this field and make the necessary preparations.