Verifiable digital twin for agricultural product traceability: a cross-modal feature alignment learning model constrained by zk-SNARK

Abstract

In the domain of agricultural product traceability, while traditional blockchain technologies ensure data immutability, they struggle to verify the authenticity of digital twins generated by generative artificial intelligence (GAI), resulting in a se mantic gap between the physical world and its virtual representation. To address these challenges, this paper proposes Verifiable Twin model driven by Cross-Modal Alignment (VTA-CMAD), which targets three core issues: cross-modal consistency verification between blockchain-stored data and AIGC-generated twins, lightweight zero-knowledge proof framework construction, and incentive-compatible suppression of malicious behaviors. The innovation of this article is reflected in three aspects. Firstly, this article proposes a 3D multimodal alignment algorithm that integrates dynamic time warping. B y integrating physical sensing temporal data, production process images, and cultural semantic descriptions, the optimal transmission mapping of feature space is established. Secondly, design a verifiable circuit zk Vector to transform the inference process of the fine-tuning diffusion model into zero knowledge proof constraints, generating proof files with a size less than 1.2KB. Finally, a dynamic consensus mechanism Proof of Trustworthiness based on Feature Alignment (PoTV) based on feature alignment is constructed to achieve adaptive adjustment of data weights. Experimental results demonstrate that the proposed approach achieves a tamper detection rate of 86.2%, a Gini coefficient of 0.19 for incentive fairness, and reduces multimodal alignment error to 0.11 ± 0.03.