1 of 10

ZKML: �When Machine Learning Meets Cryptography

2 of 10

Zero-knowledge proofs in Machine Learning

  • Introduction
  • THE scheme
  • Ezkl Library in a nutshell
  • Applications
  • Challenges and Future Directions

3 of 10

4 of 10

Introduction

THE equation:

  • oracles + neural networks + zk-SNARKs (EZKL) + smart contracts -> a framework for verifiable AI computations.

​

  • Oracles – applications that give access to (often unreliable) off-chain data

​

  • Neural networks - black boxes -> difficult to validate their outputs

​

  • Zero-knowledge proofs – transparent verification of AI without leaking sensitive info (e.g. inputs)

​

  • Smart contracts – enforceable and transparent programs that enable automated and trustless verification.

5 of 10

6 of 10

Ezkl Library

  • EZKL: ezkl - library that enables the generation of zero-knowledge proofs of complex computations.

​

  • Inputs + Parameters + Outputs -> Easily verifiable proof of computations
    • “I correctly ran this publicly available neural network on some public data and it produced this output”

​

  • Verifiable AI Computations: The library integrates with smart contracts to provide a tamper-proof and decentralized platform for verifying AI model outputs.

7 of 10

Applications

  • Using of LLMs to handle large number of queries.
    • DAOs - filtering through proposals

​

  • Image classifiers and generators can be safely used.

​

​

  • A powerful substitute for oracles that could easily be fraudulent.

8 of 10

Challenges and Future Directions

  • Scalability of zk-SNARKs: computationally complex, too costly to do on bigger scales

​

  • Efficiency of zk-SNARK generation: Improving the efficiency of zk-SNARK generation can enhance the overall performance and feasibility of the verifiable AI computations.

​

  • Governance – ensuring the spread of this technology as AI becomes more popular in blockchain.

9 of 10

10 of 10

Thank you