Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-Knowledge Proof Based Verifiable Inference of Models

Yunxiao Wang

TL;DR

The paper tackles the problem of verifying AI inference correctness without exposing proprietary parameters. It introduces a zero-knowledge framework built on PLONK/Kimchi with recursive SNARKs and no trusted setup, enabling constant-size proofs that scale with model size. By leveraging CRPC for matrix multiplication and modular components (GeMM, RMSNorm, RoPE, Softmax, SiLU, etc.), it demonstrates ZK-DeepSeek, a SNARK-verifiable version of DeepSeek-V3 with a 671B-parameter scale. The work shows practical verifiability for large LLMs and points to GPU-accelerated implementations to make verifiable AI widely feasible in real-world deployments.

Abstract

Recent advances in artificial intelligence (AI), particularly deep learning, have led to widespread adoption across various applications. Yet, a fundamental challenge persists: how can we verify the correctness of AI model inference when model owners cannot (or will not) reveal their parameters? These parameters represent enormous training costs and valuable intellectual property, making transparent verification difficult. In this paper, we introduce a zero-knowledge framework capable of verifying deep learning inference without exposing model internal parameters. Built on recursively composed zero-knowledge proofs and requiring no trusted setup, our framework supports both linear and nonlinear neural network layers, including matrix multiplication, normalization, softmax, and SiLU. Leveraging the Fiat-Shamir heuristic, we obtain a succinct non-interactive argument of knowledge (zkSNARK) with constant-size proofs. To demonstrate the practicality of our approach, we translate the DeepSeek model into a fully SNARK-verifiable version named ZK-DeepSeek and show experimentally that our framework delivers both efficiency and flexibility in real-world AI verification workloads.

Zero-Knowledge Proof Based Verifiable Inference of Models

TL;DR

The paper tackles the problem of verifying AI inference correctness without exposing proprietary parameters. It introduces a zero-knowledge framework built on PLONK/Kimchi with recursive SNARKs and no trusted setup, enabling constant-size proofs that scale with model size. By leveraging CRPC for matrix multiplication and modular components (GeMM, RMSNorm, RoPE, Softmax, SiLU, etc.), it demonstrates ZK-DeepSeek, a SNARK-verifiable version of DeepSeek-V3 with a 671B-parameter scale. The work shows practical verifiability for large LLMs and points to GPU-accelerated implementations to make verifiable AI widely feasible in real-world deployments.

Abstract

Recent advances in artificial intelligence (AI), particularly deep learning, have led to widespread adoption across various applications. Yet, a fundamental challenge persists: how can we verify the correctness of AI model inference when model owners cannot (or will not) reveal their parameters? These parameters represent enormous training costs and valuable intellectual property, making transparent verification difficult. In this paper, we introduce a zero-knowledge framework capable of verifying deep learning inference without exposing model internal parameters. Built on recursively composed zero-knowledge proofs and requiring no trusted setup, our framework supports both linear and nonlinear neural network layers, including matrix multiplication, normalization, softmax, and SiLU. Leveraging the Fiat-Shamir heuristic, we obtain a succinct non-interactive argument of knowledge (zkSNARK) with constant-size proofs. To demonstrate the practicality of our approach, we translate the DeepSeek model into a fully SNARK-verifiable version named ZK-DeepSeek and show experimentally that our framework delivers both efficiency and flexibility in real-world AI verification workloads.

Paper Structure

This paper contains 25 sections, 29 equations, 12 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Notations
  4. PLONK and Kimchi
  5. Pickles
  6. Constraint-Reduced Polynomial Circuits
  7. DeepSeek
  8. Methodology
  9. General-purpose Matrix Multiplication (GeMM)
  10. Root Mean Square Layer Normalization (RMSNorm)
  11. Embedding
  12. Rotary Positional Embedding (RoPE)
  13. Softmax
  14. Sigmoid and Sigmoid Linear Unit(SiLU)
  15. Sort and top-k selection
  16. ...and 10 more sections

Figures (12)

  • Figure 1: The multiplication in CRPC
  • Figure 2: The proof construction of matrix X
  • Figure 3: The proof construction of matrix W
  • Figure 4: The proof composition of the multiplication of X and W
  • Figure 5: The row proof composition of the RMSNorm
  • ...and 7 more figures