Guo et al. propose a privacy-preserving zkML scheme based on architecture-private zero-knowledge proof, using parameterized R1CS and functional relationship proofs to hide CNN model architecture, achieving 30% slower proof time than BFG+23 on VGG16. Guo等人在论文中提出了一种架构隐私的zkML方案,通过参数化R1CS和功能关系证明来隐藏CNN模型架构,在VGG16上仅比BFG+23慢30%。
Notes
Existing zkML schemes mainly hide model parameters but expose the complete CNN architecture information.
Proposes parameterized R1CS (pR1CS) as a generalization of R1CS to allow the prover to submit the model architecture.
Introduces a functional relationship proof scheme to verify that the submitted architecture is valid.
When batch proving 64 instances on the VGG16 model, proof time is only 30% slower than BFG+23.
Using pR1CS to prove matrix multiplication is at least 3 times faster than traditional R1CS.
This scheme provides dual privacy protection for both the architecture and parameters of neural networks in zero-knowledge proofs.