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

作　　者：崔思颖 谭志杰 袁想 李伟平[1] 莫同[1] 乔秀全[2] 吴中海[1] Cui Siying;Tan Zhijie;Yuan Xiang;Li Weiping;Mo Tong;Qiao Xiuquan;Wu Zhonghai(School of Software and Microelectronics,Peking University,Beijing 102600,China;State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京大学软件与微电子学院,北京102600 [2]北京邮电大学网络与交换技术国家重点实验室,北京100876

出　　处：《南京师大学报(自然科学版)》2025年第2期102-111,共10页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：辽宁省科学技术计划揭榜挂帅项目(2021JH1/10400010).

摘　　要：如今,区块链技术被应用到包含电子证照、人脸图像等政府数据共享领域,但当前的大型区块链系统普遍面临低带宽和高存储成本的问题.本文提出了一种适用于政务区块链的跨模态人脸生成模型,将人脸图像转换为文本模态存储在链上,用户可使用文本与掩膜生成指定人的人脸图像.首先利用多任务学习方法训练基于ResNet-18网络结构的人脸分类器,将人脸图像转换为身份代号文本存储在链上.然后设计了区域感知码本和基于Transformer结构的混合专家采样器,采样器采用扩散模型的方法从码本中采样索引,采样结果由一个可学习的解码器转换成细粒度的人脸图像.在进行数据增强后的Casia Face V5数据集上的实验表明,模型在人脸分类任务中准确率可达95%以上,压缩效果达到了传统图像压缩方法1/10000的持久化时间与1/200的文件大小,与其他先进人脸图像生成方法相比,此模型可以可控地生成高保真度的指定人的人脸图像,并以1/20的参数量达到与大型预训练模型相近的人脸生成效果.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 1/10000 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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]