ZKP学习笔记

ZK-Learning MOOC课程笔记

Lecture 14.2 Other ZKP Applications (Yupeng Zhang)

ZKP for Machine Learning
- Zero-knowledge proof without revealing the ML models
  - Fairness of ML models
  - Integrity of ML inferences (The decision is made by the ML model)
- Challenges
  - Efficiency and Scalability of general-purpose SNARKs
    - scale to <230 = 1 billion gates (64 GB RAM), prover time minutes to hours
  - VGG 16 on CIFAR-10 15 million parameters in the model 1.1 billion gates for an inference
- Solution: Special-Purpose ZKPs
  - ZKP for Matrix Multiplication [Thaler’13]
  - ZKP for 2-D Convolutions [LXZ’21]
- Other Related Works on ZKML
  - ZKDT[ZFZD20], vCNN [LKKO20], ZEN [FQZ+21], Mystique [WYX+21], pvCNN [WWT+22], [KHSS22], …
ZKP for Program Analysis
- Zero-knowledge Program Analysis
  - secret program P
  - public function: static analysis algorithm
  - Prove the safety properties of P
- Zero-knowledge Vulnerability Disclosure
  - secret vulnerability
  - public program
  - Running the program leads to crash
- Challenges
  - ZKP schemes support circuits.
  - Program analysis is usually RAM computation
- Solution: Auxiliary Inputs
  - Ask the prover to provide additional data as the input of ZKP
    - Not trusted
    - Not sent to the verifier
    - Significantly improves the efficiency of ZKP
  - Example
    - Prover provides the final state of the list
    - Prover provides head and tail of each step
    - The circuit checks the correctness (offline memory checking [BEGKN’91,Setty’20, …])
- Related works
  - Static analysis: [FDNZ’21, LAHPTW’22, …]
  - Vulnerabilities: [GHHKPV’22, CHPPT’23, …]
ZKP for Middlebox
- Middleboxes inspect traffic to ensure security policy
- Zero-Knowledge Middleboxes [GAZBW’22]
- Challenges
  - Work with TLS 1.3
  - Legacy cryptographic functions such as AES, SHA