本项目旨在研发识别精度高、安全性能好、抗干扰能力强的大规模3D人脸识别系统。聚焦深度学习框架下大规模3D人脸识别所面临的数据匮乏、特征表达及识别、3D面具防伪、对抗样本攻击及防御四个一体化技术瓶颈问题。旨在解决上述瓶颈背后的五个关键科学问题：1.深度编码-解码网络学习框架的大规模非线性判别式3D人脸重建方法；2.基于深度融合、迁移、协同学习的2D辅助大规模3D人脸识别方法；3.基于几何深度学习的3D人脸形状特征学习匹配及识别方法；4.基于深度学习的面部时空模态多属性特征联合学习的3D面具防御策略；5.基于最优化问题扰动分析理论的3D人脸对抗样本攻击及防御机理。项目研究内容紧密相关、逐层递进、共同围绕3D人脸识别系统的识别精度、安全性和鲁棒性三个核心要素。本项目属于典型的“需求牵引，突破瓶颈”类型科学问题。项目的顺利实施将为我国3D人脸识别技术迈向实际应用奠定理论和技术研究基础。

The aim of this proposal is to develop high-accuracy, high-security, and high-robustness large scale three-dimensional (3D) face recognition system. In particularly, we are focusing on the following bottleneck problems under the framework of deep learning-based large-scale 3D face recognition: the lack of 3D face data, learning representation for 3D face and improving the accuracy of 3D face recognition techniques, 3D mask face anti-spoofing, attacking and defensing 3D face adversarial samples. We are aiming to solve the core scientific issues behind the technical bottlenecks mentioned above: 1. Discriminative and non-linearly large-scale 3D face reconstruction methods based on deep auto-encoder and decoder neural network; 2. 2D-aided large-scale 3D face recognition methods based on deep fusion learning, deep transfer learning, and deep co-training; 3. Geometric deep learning based 3D face shape feature learning, matching and recognition methods; 4. 3D mask face anti-spoofing techniques based on deep joint learning the temporal-spatial and modality multiple facial attributes; 5. Attacking and defensing adversarial samples based on perturbation analysis theory of the optimization. All the research topics are closely related, progressive-like and focusing on the three core factors of accuracy, security, and robustness of 3D face recognition system. This project belongs to a typical scientific problem of "demand pull, break through bottleneck" type. Successfully running of this proposal will establish the theory and technique basis for the real application of 3D face recognition techniques in our country.