南京师大学报（自然科学版）2025，Vol.48Issue(2)：102-111,10.DOI:10.3969/j.issn.1001-4616.2025.02.011
一种适用于政务区块链的跨模态人脸生成模型
Cross-Modal Face Generation Model for Government Blockchain
摘要
Abstract
Blockchain technology is currently used in government data sharing,but faces challenges such as limited bandwidth and high storage costs.To address this,the study proposed a cross-modal face generation model for the government blockchain.This model converted face images into text modals and stored them on the chain,allowing users to generate face images of specific individuals using text and masks.To achieve this,the study trained a face classifier based on the ResNet-18 network structure using a multi-task learning method.The resulting identity code text is then stored on the blockchain.Additionally,the study constructed region-aware codebooks and designed a diffusion-based transformer sampler with mixture-of-experts.This sampler converts indexed from the codebooks into fine-grained face images using a learnable decoder.The experiments on the enhanced Casia Face V5 dataset demonstrated that the model achieved a face classification accuracy rate of 95%.Furthermore,it offered a persistence time of 1/10 000 and a file size of 1/200 compared to traditional image compression methods.Compared to other advanced face image generation methods,this model can generate high-fidelity face images of specific individuals while requiring only 1/20 of the parameters of large pre-trained models.关键词
区块链/跨模态人脸生成/可控图像生成/扩散模型/人脸识别Key words
blockchain/cross-modal face generation/controllable image generation/diffusion model/face recognition分类
计算机与自动化引用本文复制引用
崔思颖,谭志杰,袁想,李伟平,莫同,乔秀全,吴中海..一种适用于政务区块链的跨模态人脸生成模型[J].南京师大学报（自然科学版）,2025,48(2):102-111,10.基金项目
辽宁省科学技术计划揭榜挂帅项目(2021JH1/10400010). （2021JH1/10400010）
