融合深度学习与特征提取的指纹伪造检测方法研究

Fingerprint-based authentication has been applied in various applications such as border identity authentication, attendance system in enterprises, and authorization of payment in smartphones. However, fingerprint-based authentication suffers from security risks. First of all, human fingerprints are likely to be stolen, and the spoof fingerprints can be made with cheap materials to achieve illegal authentication. Besides, people may make spoof fingerprints for themselves to deceive the attendance system. Therefore, it is of great practical value to study how to distinguish the real and spoof fingerprints against illegal authentication. The existing methods utilized the high-dimensional feature vectors or deep learning to detect the spoof fingerprints. However, they were not specially-designed for the fingerprint spoof detection problem and got unsatisfied results in the detection of spoof fingerprints made by the unknown materials. Accordingly, our project plans to design new feature extraction methods and deep neural network models to improve the detection accuracy of spoof fingerprints made by the unknown materials. The main research contents are organized as follows. Firstly, the small structures of real fingerprints are difficult to copy. Thus, the morphologic features of small structures are studied. Secondly, the real and spoof fingerprints hold different elasticity. Accordingly, the deformation of the fingerprint is detected as the features. Thirdly, the texture features are also studied to be used together with the morphologic features and elasticity features. Finally, how to construct the convolution neural network that is suitable for the fingerprint spoof detection is studied. The research of this project will provide new methods for fingerprint spoof detection. It is expected to increase the detection accuracy to 96% in the material-cross detection of spoof fingerprints.

指纹认证已广泛应用到国家安全和社会生活的各个领域。然而人类指纹易被窃取，攻击者可通过伪造指纹来实现非法认证。因此，研究指纹伪造检测算法来检测伪造指纹，具有很强的安全需求与社会价值。现有研究运用高维特征或深度学习来检测伪造指纹，算法泛化能力不强，跨材料检测精度不高。本课题将针对指纹伪造检测问题的特点，研究设计新的特征提取方法和深度学习模型，旨在提高算法的跨材料检测能力。主要内容包括：①由于伪造指纹难以复制真指纹的微小形态结构，研究面向跨材料检测的指纹形态特征提取方法；②根据手指皮肤与伪造指纹材料在弹性方面的差异，研究面向跨材料检测的指纹弹性特征提取方法；③研究面向指纹伪造检测的纹理特征提取方法，与形态特征和弹性特征形成互补；④针对指纹伪造检测问题的特点，研究面向指纹伪造检测的深度学习模型设计方法。课题的研究将为指纹伪造检测提供新方法，力争将算法跨材料检测的平均精度提高到96%以上。

伪造指纹给指纹识别和认证系统的安全性带来了巨大挑战，攻击者可通过伪造指纹来实现非法认证行为，因此研究伪造指纹检测具有重要的社会安全价值。然而，现有研究直接运用高维特征或深度学习方法来检测伪造指纹，算法泛化能力不强，跨材料检测精度不高，且检测模型复杂度高，效率低。针对以上缺陷，本课题结合数字取证，隐私保护，深度学习等方面算法和技术，设计了多种高效、高泛化的伪造检测特征提取方案以及深度学习模型构建方案。在面向跨材料检测的形态特征提取方法研究方面，提出了基于空域脊线形态一致性的指纹伪造识别方案，通过探索指纹脊线的空间结构关系，在跨材料和跨传感器条件下均实现了极高的检测精度；在面向跨材料检测的弹性特征提取方法研究方面，提出了基于多模态特征融合的指纹活体检测方案，利用加权多模态卷积神经网络提取并融合不同的深度特征，达到了更加稳健的检测效果；在面向指纹活性检测的纹理特征提取方法研究方面，提出了基于均匀局部二值模式纹理的指纹活体检测方案，通过均匀局部二值模式描述符提取指纹图像中的纹理特征，并利用高效的增量式浅层神经网络进行训练，实现了精度不变，但规模更小、计算复杂度更低的检测方案；在面向指纹伪造检测的深度学习模型设计研究方面，提出了一种轻量化的多尺度卷积神经网络伪造指纹检测模型，利用混合空间金字塔池化，极大地减少了模型参数量，采用注意力机制进行优化，缓解了模型过拟合的问题，提高了检测模型的泛化能力。项目研究期间，项目负责人共发表学术论文19篇，其中SCI论文15篇，包括IEEE汇刊论文5篇，申请和授权发明专利11项，培养研究生7人。本项目的研究成果对生物特征保护乃至数据隐私保护领域都有着重要贡献。