Blockchain technology and machine learning, as two highly regarded fields, are leading technological advances with their decentralized nature and data-driven capabilities. ZK (Zero-Knowledge) is a concept in cryptography within blockchain technology, referring to a proof or interactive process where 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 to form models, and can make predictions and decisions.
In this context, the combination of ZK and ML, known as ZKML (Zero-Knowledge Machine Learning), is thriving. ZKML combines the privacy protection and verification capabilities of zero-knowledge proofs with the data processing and decision-making capabilities of machine learning, bringing new opportunities and possibilities to blockchain applications. ZKML provides us with a solution that simultaneously protects data privacy, verifies the accuracy of models, and improves computational efficiency.
This article will delve into ZKML, explore its technical principles and application scenarios, and unravel how ZKML builds a more comprehensive, secure, and efficient digital future with enhanced privacy.
ZKML: The Combination of Zero-Knowledge Proofs and Machine Learning
There are two reasons why zero-knowledge proofs and machine learning can be combined on the blockchain:
On the one hand, the developers of ZK not only aim to achieve efficient verification of on-chain transactions but also hope to apply ZK in a broader ecological field. The powerful support of ML's AI capabilities serves as an excellent enabler for the expansion of the ZK application ecosystem.
On the other hand, the entire process from the development to the usage of ML models faces the challenge of trust proof. ZK can help ML achieve effective proof of validity without leaking data and information, thus addressing the trust dilemma of ML. The combination of ZKML allows the two to meet their respective needs and adds momentum to the blockchain ecosystem.
Complementary Development Needs and Abilities of ZK and ML
ML has a lot of trust issues to solve, and the accuracy, completeness, and privacy of various workflows need to be proven. ZK can effectively verify the correctness of any type of computation while ensuring privacy, which solves the long-standing trust proof problem in machine learning. The integrity of the model is an important trust proof problem in ML training, but privacy protection of the data and information used in ML model training and usage is equally important. This makes it difficult for ML training to pass the trust proof of third-party auditing regulatory agencies. Decentralized zero-knowledge (ZK) with zero-knowledge properties is a highly compatible trust proof path for ML.
"AI enhances productivity, blockchain optimizes production relationships." ML injects higher innovative momentum and service quality into the ZK track, while ZK provides verifiability and privacy protection to ML. ZKML complements each other in a blockchain environment.
ZKML's main technical advantages combine computational integrity, privacy protection, and heuristic optimization. From a privacy perspective, ZKML's advantages are:
Transparent verification:
Zero-knowledge proof (ZK) allows evaluation of model performance without revealing internal details, enabling transparent and trustless evaluation processes.
Data privacy protection:
ZK can be used to verify public data using public models or verify private data using private models, ensuring data privacy and sensitivity.
ZK itself ensures the correctness of a statement while guaranteeing privacy through cryptographic protocols, effectively addressing the flaws of machine learning in privacy protection in the areas of privacy-protecting machine learning and homomorphic encryption machine learning. By integrating ZK into the ML process, a secure and privacy-protecting platform is created, addressing the limitations of traditional machine learning. This not only encourages privacy companies to adopt machine learning technology but also motivates Web2 developers to explore the potential of Web3 technology.
ZK Empowering ML: Providing On-Chain Infrastructure
One reason why ML, which has already matured off-chain, has just entered the on-chain space is the high computational cost of blockchain. Many machine learning projects cannot directly run in blockchain environments represented by EVM due to computational limitations. Additionally, although ZK's validity verification is more efficient than duplicate computation, this advantage is limited to transaction data processing native to blockchain. When ZK's already complex cryptographic operations and interactions face a large number of computations in ML, the low TPS problem of the blockchain is exposed, becoming the biggest obstacle to ML on-chain.
ZK-SNARKs have alleviated the high computational power requirements of ML. ZK-SNARKs, short for "Zero-Knowledge Succinct Non-Interactive Argument of Knowledge," is a cryptographic construction of zero-knowledge proofs. It is based on elliptic curve cryptography and homomorphic encryption and is used to achieve efficient zero-knowledge proofs. ZK-SNARKs have the characteristic of being highly compact. With the use of ZK-SNARKs, a prover can generate a short and compact proof, while a verifier only needs to perform a small amount of computation to verify the validity of the proof, without requiring multiple interactions with the prover. This property of only requiring a single interaction between the prover and verifier makes ZK-SNARKs efficient and practical in real-world applications, making them well-suited for the on-chain computational power requirements of ML. Currently, ZK-SNARKs are the main form of ZK in ZKML.
On-chain infrastructure requirements for ML and corresponding projects
The empowerment of ML by ZK mainly manifests in the zero-knowledge proofs throughout the entire process of ML, which involves the interaction between ML and on-chain functionality. The two main problems that need to be addressed in this interaction are aligning the data formats of both ML and the chain and providing computational power for the ZK proof process.
Hardware acceleration for ZK: ML ZK proofs are relatively complex, requiring hardware-assisted on-chain computational power for proof calculation acceleration. Projects in this category include: Cysic, Ulvetanna, Ingonyama, Supranational, Accseal.
On-chain data processing for ML: Processing on-chain data into a format that can be used for ML training and facilitating easy access to ML's output results from the chain. Projects in this category include: Axiom, Herodotus, LAGRANGE, Hyper Oracle.
ML circuitization: ML computation patterns and on-chain circuitized ZK proofs have some differences, and ML's on-chain operations must transform its computation patterns into circuit forms that can be processed by blockchain ZK. Projects in this category include: Modulus Labs, Jason Morton, Giza.
ZK proof of ML results: Trust verification of ML requires ZK on the chain. Applications constructed based on ZK-SNARKs on Risc Zero or Nil Foundation can achieve proof of model authenticity. Projects in this category include: RISC Zero, Axiom, Herodotus, Delphinus Lab, Hyper Oracle, Poseidon ZKP, IronMill.
ML Empowering ZK: Enriching Web3 Application Scenarios
ZK solves the trust issue of ML and provides on-chain opportunities for ML. Many Web3 domains urgently need the productivity or decision support of AI ML. ZKML enables on-chain applications to empower AI while ensuring decentralization and effectiveness.
DeFi
ZKML can help automate DeFi, firstly by automating the update of on-chain protocol parameters, and secondly by automating trading strategies.
Modulus Labs has introduced RockyBot, the first fully on-chain AI trading robot in history.
DID
ZKML can assist in the development of decentralized identity (DID) in Web3. Previously, identity management models such as private keys and mnemonic phrases have resulted in a poor user experience in Web3. True DID development can be achieved through ZKML by identifying the biometric information of Web3 entities. At the same time, ZKML ensures the security of user biometric information privacy.
Worldcoin is applying ZKML to achieve zero-knowledge DID verification based on iris scans.
Games
ZKML can help Web3 games achieve full functionality on-chain. ML can bring differentiated automation to game interactions, increasing the fun of games, while ZK can enable on-chain interaction decisions of ML.
Modulus Labs has introduced @VsLeela, a chess game driven by ZKML;
AI ARENA has utilized ZKML to achieve high interactivity in on-chain NFT games.
Healthcare and Legal Consulting
Healthcare and legal consulting are fields with high privacy requirements and the need for extensive case accumulation. ZKML can help users make decisions while ensuring the privacy of users is not compromised.
Challenges Faced by ZKML
ZKML is currently experiencing rapid development, but it faces two main challenges in the future due to not being native to blockchain and requiring a significant amount of computing power:
Parameter distortion issue in the process of quantifying ML data on-chain:
Most ML models use floating-point numbers to represent their parameters, while ZK circuits require the use of fixed-point numbers. In this process of converting number types, the precision of ML parameters will be reduced, which to some extent leads to distortion in the output results of ML.
The problem of high computational power required for ZK proofs of large-scale models:
Currently, the computational power of blockchain is unable to handle large-scale and highly computational ZKML on-chain. The popular ZK-SNARKs only support small-scale and less computational ML zero-knowledge proofs. The limited computational power is a critical factor affecting the development of ZKML blockchain applications.
The computation complexity of the ZK proof generation stage is high and requires a large amount of computational resources. Due to the high correlation between the data typically accessed and processed during the ZK proof generation stage, it is difficult to distribute this process, making it "non-parallelizable". Distributing this process may introduce additional complexity and even reduce overall performance. Currently, to address the efficiency of ZK computations, the mainstream research direction focuses more on algorithm optimization and hardware acceleration.
Conclusion
ZKML is the convergence of zero-knowledge proofs and machine learning, and the evolving blockchain technology ZK helps solve trust proof problems for ML and provides an on-chain environment for ML. Mature AI technology ML assists ZK in achieving Web3 ecosystem expansion and application innovation.
The development of ZKML faces some challenges, such as the parameter distortion problem and the high computational power requirements of large-scale models, but these problems can be addressed through technological innovation and hardware acceleration. With the emergence and development of ZKML projects, we can foresee that it will bring more innovation and value to the Web3 ecosystem in areas such as DeFi, DID, gaming, healthcare, etc.
In the future, ZKML is expected to become the key to unlocking the integration of Web3 and AI, providing strong support for building secure, privacy-protected, and efficient blockchain applications. By combining the zero-knowledge nature of ZK with the data processing capabilities of ML, we can definitely create a more open, intelligent, and trusted digital world!
