CNN Verification via Zero-Knowledge Proof Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) are vulnerable to malfunctions and errors, particularly in critical applications like healthcare and legal systems, where incorrect predictions can have severe consequences, necessitating a robust verification method to ensure accuracy and trustworthiness.
Innovation Solution
A method and device for verifying CNN models by separating convolution operations from other operations, generating commitment and proof values using zero-knowledge proof schemes, specifically employing polynomial circuit equations and arithmetic circuit equations to reduce verification time and ensure public verification while protecting privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If zero-knowledge proof is applied to verify CNN model, then verification reliability is improved, but verification time increases significantly
Solution Approach 1:
The patent divides the CNN verification process into separate modules: generating commitment values for convolution operations and other operations independently, then combining them through connection proofs. This segmentation allows parallel processing and reduces the computational burden on the verifier, significantly decreasing verification time while maintaining reliability.
Solution Approach 2:
The patent introduces commitment values as intermediary elements that bridge the convolution operation verification and other operation verification. By using commitment values and connection proofs as mediators, the system enables efficient verification without requiring direct computation of all neural network operations, thus reducing verification time.
2Device complexity
If convolution operation and other operations are processed together, then device complexity is reduced, but verification time increases
Solution Approach 1:
The patent segments the verification process into distinct stages: first generating commitment values for convolution operations, then generating commitment values for other operations, and finally creating connection proofs to link them. This segmentation maintains manageable system complexity while enabling time-efficient verification through structured processing.
Solution Approach 2:
The patent performs preliminary actions by generating commitment values before the actual verification process. By pre-computing and storing commitment values for both convolution and other operations, the system reduces the computational burden during verification, thereby decreasing verification time without increasing complexity.
3Ease of operation
If public verification is enabled, then verification transparency is improved, but privacy protection of neural network parameters deteriorates
Solution Approach 1:
The patent creates cryptographic copies of neural network parameters in the form of commitment values. These commitments allow public verification of model correctness without revealing the actual parameter values. The commitment values serve as encrypted representations that can be verified publicly while preserving parameter privacy.
Solution Approach 2:
The patent uses commitment values and connection proofs as intermediaries between the neural network parameters and the verification process. These intermediaries enable transparent verification of computation correctness while preventing direct exposure of sensitive parameter information, thus maintaining both accessibility and privacy.
Data Source
AI summary
Disclosed are a method for verifying a convolutional neural network model and a device thereof. The method for verifying the convolutional neural network model includes (a) generating a polynomial circuit equation for a first configuration of a plurality of configurations configuring the convolutional neural network model; (b) generating a first commitment value and a first proof value by applying a zero-knowledge proof scheme based on the polynomial circuit equation; (c) generating an arithmetic circuit equation for a second configuration of the plurality of configurations; (d) generating a second commitment value and a second proof value by applying a zero-knowledge proof scheme based on the arithmetic circuit equation; and (e) generating a connection proof value connecting the first commitment value and the second commitment value.