The blockchain interoperability infrastructure Polyhedra, after experiencing a drop in coin prices and a setback in the competition for the $ZK code with zkSync, has recently revitalized its efforts by launching 'the universal chain built for AI', known as EXPchain. It proposes the concept of Proof of Intelligence (PoI), creating a tamper-proof, trustworthy blockchain for AI models. Whether the combination of zk and AI will be a successful transformation remains to be seen.

Traditional AI regulation involves sensitive data, zkML offers a new solution

The official definition of EXPchain is a blockchain protocol designed specifically for scalable, verifiable, and privacy-focused AI applications. As 'the universal chain built for AI', EXPchain integrates zero-knowledge machine learning (zkML) and a new proof of intelligence (PoI) framework. Major innovations include the efficient zk proving system Expander, as well as the developer-friendly zkPyTorch toolkit that integrates zkML into traditional AI workflows.

Artificial intelligence plays an increasingly critical role in various industries, from using facial recognition to unlock smartphones to AI-driven loan applications and medical diagnoses. These technologies bring both tremendous potential and challenges, such as how to ensure that AI systems operate fairly, accurately, and safely? How to protect sensitive data without compromising transparency and accountability?

Additionally, various governments are beginning to regulate AI, such as the EU's AI Act and the U.S. National Institute of Standards and Technology (NIST) AI Risk Management Framework. The problem with traditional approaches is that they require disclosing proprietary models or sensitive data, leading to trade-offs between security, privacy, and trust.

Zero-knowledge machine learning (zkML) offers a solution different from traditional methods. The unique features of zero-knowledge proofs can achieve mathematical verification of AI systems while protecting data and model privacy. Polyhedra has launched the interoperability protocol EXPchain based on zkML technology, which not only balances AI behavior and compliance regulations but also provides scalable and secure verification.

Technical debt continues to expand, on-chain AI transaction processes benefit accountability

A study indicates that in 2022, the software technology debt in the United States (which refers to compromises made during software development to quickly launch or meet short-term demands, often increasing long-term system maintenance costs) has grown to $2.41 trillion. Additionally, research from one of the world's four major consulting firms, PricewaterhouseCoopers (PwC), also indicates that by 2030, AI is expected to contribute up to $15.7 trillion to the global economy.

As the scale of AI expands, it may exacerbate the expansion of technical debt. In this regard, the business column Raconteur has questioned whether companies are prepared to bear the costs of AI failures. AI failures include incorrect output, data breaches, and cyber attacks. In addition to economic losses, these mistakes often cause harm to individuals.

For example, incorrect output data may lead to machine misjudgments or biased decisions. Therefore, ensuring that every element driving AI-powered transactions is verifiable and accountable, from data input to model output, is essential. Addressing these risks is crucial to unleash the full potential of artificial intelligence. This is where the EXPchain real-time verification blockchain comes into play.

Three major technological innovations: Can Polyhedra solve the zk prover generator problem?

Technological innovations include Expander, ExPos, and zkPyTorch

Polyhedra: Expander is currently the world's fastest zk prover

Data provided by Polyhedra includes:

  • Processing VGG-16 images on a single-threaded CPU takes only 2.2 seconds

  • Processing Llama-3.1 8B on a single-threaded CPU takes 150 seconds per token

  • Performance is four orders of magnitude faster than previous data

These advancements significantly reduce the cost and delay of AI verification, supporting various applications from privacy inference to model auditing. Expander also aligns with the zk end-state vision proposed by Vitalik Buterin.

Layer 2 is mainly divided into Optimistic Rollup and zk Rollup. For most zk Rollup public chains, ZKP proving generation is a bottleneck, and companies must deploy powerful machines with TB memory to handle the large number of transactions in ZKP. Previously, Polyhedra's CTO Tiancheng Xie and Chief Scientist Jiaheng Zhang's team proposed a paper discussing new solutions using fully decentralized ZKP to improve the scalability of zk technology.

ExPoS: extended proof of stake

ExPoS is a proof of stake mechanism developed for the zkML technology in EXPchain, which verifies the behavior and compliance of AI applications without leaking proprietary model data. In layman's terms, it uses Polyhedra's zkBridge technology to unify and connect all staking mechanisms on the blockchain into a cohesive staking network.

zkPyTorch: a developer-friendly toolkit

zkPyTorch automatically converts PyTorch operations into zk circuits, bridging the gap between traditional AI development workflows and zero-knowledge machine learning (zkML). This integration allows developers to use familiar tools while significantly reducing the time and complexity of deploying zk-supported AI applications.

zkML can complete LLM verification under privacy constraints

The core of EXPchain lies in zero-knowledge machine learning (zkML), which supports encrypted verification of AI models, achieving security and accuracy throughout the entire machine learning lifecycle, including:

  • Verifiable inference: proving AI outputs without exposing the model or data.

  • Model auditing: verifying fairness and compliance of performance based on the test set.

  • Training validation: ensuring compliance with protocols without leaking sensitive input.

zkML specific applications include:

  • Adding digital watermarks to large language models (LLMs). A digital watermark is a small, subtle feature embedded in the text generated by the LLM for identifying whether the text was generated by a specific model, used to prevent counterfeit content and content abuse.

  • Ensuring model compliance, such as compliance verification in financial institutions.

  • Achieving secure multi-party computation in privacy-focused industries.

Currently, the zkML digital watermark of EXPchain can already be used to verify large language models such as Llama-3.1 8B.

Polyhedra's chief cryptographer has a significant background, driving the PoI AI proof chain

EXPchain can be seen as Proof of Intelligence (PoI), creating a tamper-proof, trustworthy blockchain for AI models, verifying their source, authenticity, and ethical compliance. This framework protects intellectual property rights and ensures transparent accountability, linking each AI model's source and performance to verifiable on-chain records, providing unprecedented transparency for AI-driven ecosystems.

Talking about all these behind-the-scenes drivers, one cannot overlook Polyhedra's chief cryptographer, Zhenfei Zhang. He previously worked at industry leaders such as Algorand, Espresso, Ethereum Foundation, and Scroll, enjoying considerable recognition in the cryptography community. The article 'ZEN: An Optimized Compiler for Verifiable Zero-Knowledge Neural Network Inference' discusses verifiable machine learning.

This article discusses how Polyhedra launches EXPchain for AI applications, analyzing the necessity of putting AI models on-chain and the decentralized zk prover, first appearing in Chain News ABMedia.