# zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training

**Authors:** Haochen Sun, Hongyang Zhang  
**arXiv:** [2307.16273](https://arxiv.org/abs/2307.16273)  
**Date:** 2023-07-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.16273v1/

## Key points

- Authors Haochen Sun and Hongyang Zhang address concerns about the legitimacy of deep learning training processes in their paper "zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training."
- The rapid advancements in deep learning have raised questions about protecting intellectual properties of untrusted AI developers.
- zkDL is a novel zero-knowledge proof system specifically designed for deep learning training, introduced by Sun and Zhang.
- At the core of zkDL is zkReLU, a specialized zero-knowledge proof protocol optimized for proving time and proof size when dealing with the ReLU activation function.
- By integrating zkReLU into the overall proof system, Sun and Zhang devise an innovative approach leveraging parallel computation resources to reduce proving time and proof sizes significantly.
- zkDL enables efficient generation of complete and sound proofs in less than a minute with a compact size of less than 20 kB per training step.
- The system demonstrates high efficiency on a 16-layer neural network with 200 million parameters while ensuring data privacy and protecting model parameters.

## Summary

In their paper titled "zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training," authors Haochen Sun and Hongyang Zhang address concerns surrounding the legitimacy of deep learning training processes in light of recent advancements in the field. The rapid developments in deep learning have led to significant changes in various aspects of society, but they have also raised questions about the protection of intellectual properties belonging to untrusted AI developers. The challenge lies in verifying the training process without compromising sensitive model parameters and training data. To tackle this issue, Sun and Zhang introduce zkDL, a novel zero-knowledge proof system designed specifically for deep learning training. At the core of zkDL is zkReLU, a specialized zero-knowledge proof protocol optimized for proving time and proof size when dealing with the ReLU activation function—a key component that poses challenges due to its non-arithmetic nature. By integrating zkReLU into the overall proof system for training processes, the authors devise a unique construction of an arithmetic circuit derived from neural networks. This innovative approach leverages parallel computation resources to significantly reduce proving time and proof sizes by a factor corresponding to the network depth. As a result, zkDL enables the generation of complete and sound proofs efficiently, taking less than a minute with a compact size of less than 20 kB per training step. This level of efficiency is demonstrated on a 16-layer neural network with 200 million parameters while ensuring data privacy and protecting model parameters. Overall, Sun and Zhang's zkDL system offers a promising solution to address concerns regarding deep learning training processes by providing efficient zero-knowledge proofs that safeguard intellectual properties and maintain data privacy throughout the verification process.