结合自注意力特征融合的人脸模板重建技术  

Face Template Reconstruction Technology Combined With Self-Attention Feature Fusion

作　　者：王亚男 董建民[1] 孙冰阳 王宁博 WANG Yanan;DONG Jianmin;SUN Bingyang;WANG Ningbo(College of Information Engineering,Xizang Minzu University,Xianyang 712082,China)

机构地区：[1]西藏民族大学信息工程学院,陕西咸阳712082

出　　处：《软件导刊》2025年第3期177-184,共8页Software Guide

基　　金：西藏自治区重点研发计划(XZ202401JD0009);西藏民族大学校内科研基金项目(22MDY013)。

摘　　要：针对人脸识别系统模板反演重建方法存在的人脸模板特征数据单一,模板向量不足以捕捉多样性,生成的图像在细节上不够丰富的问题,提出一个新的人脸模板映射网络(FFMapNet)。首先,通过自注意力特征提取模块(SAFEM),结合反卷积层对初始人脸模版特征进行上采样,并引入高维通道特征;其次,利用自注意力机制的全局信息建模功能,从模板特征反演中提取并增强特征;最后,利用特征融合模块(FFM)进一步处理和融合增强后的特征,提升人脸模板到人脸生成器网络中间潜在空间(LS)映射的一致性,生成更具表达能力的特征映射。在MOBIO、LFW和AgeDB数据集上对最先进的人脸识别系统进行模板反演攻击,实验结果表明所提方法的重建质量与攻击成功率均优于现有方法。Aiming at the problem that the face template feature data of the template inversion and reconstruction method in face recognition system is single,the template vector is not enough to capture diversity,and the generated image is not rich in details,a new face template mapping network(FFMapNet)is proposed.Firstly,the self attention feature extraction module(SAFEM)is used in conjunction with the deconvolution layer to upsample the initial facial template features and introduce high-dimensional channel features;Secondly,utilizing the global information modeling function of self attention mechanism,features are extracted and enhanced from template feature inversion;Finally,the feature fusion module(FFM)is used to further process and fuse the enhanced features,improving the consistency of the latent space(LS)mapping from the face template to the face generator network,and generating more expressive feature maps.Template inversion attacks were conducted on state-of-the-art facial recognition systems on MOBIO,LFW,and AgeDB datasets,and experimental results showed that the proposed method outperformed existing methods in terms of reconstruction quality and attack success rate.

关 键 词：人脸识别 模板反演 黑盒攻击 自注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]