zkPoT: ZKP based on machine learning model training

1 Introduction

In the 2023 paper Experimenting with Zero-Knowledge Proofs of Training by Sanjam Garg et al ., the zkPoT (zero-knowledge proof of training) protocol designed:

• For streaming-friendly.
• The RAM required is not proportional to the size of the training circuit.
• Combined MPC-in-the-head + zk-SNARKs.
• The total proof size is 10% less than the training data set size and is divided into 3 stages: [Take training a logistic regression model using mini-batch gradient descent on a 4 GB dataset of 262,144 records with 1024 features as an example]
  • Data independent offline stage.
  • Depends on data, but does not rely on the model stage. This stage:
    • Prover takes about 1 hour.
    • Verifier takes a few seconds.
  • Online stage: relies on both data and models. Online stage:
    • Prover takes less than 10 minutes.
    • Verifier takes less than half a minute.
      

For open source code implementation, see:

• https://github.com/guruvamsi-policharla/zkpot（Rust）