基于频域解离特征的OCT指纹表征攻击检测

OCT Fingerprint Presentation Attack Detection Using Frequency Disentangling Features

在自动指纹识别系统中,指纹防伪能力的发展至关重要.传统指纹一般由表面成像得到,而这种表面的纹理特征极容易被盗取.基于这种传统指纹的识别系统,检测指纹表征攻击的能力十分有限.因此,现有研究普遍针对具有防伪特征的指纹模态,如具有汗腺特征的高精度指纹和具有指静脉特征的指纹开发表征攻击检测算法.在本篇工作中,为了进一步提高指纹系统的表征攻击检测能力,我们提出一种基于光学相干断层扫描技术(Optical Coherence Tomography,OCT)的频域指纹表征攻击检测方法.与以往方法不同,我们首先利用卷积神经网络和残差结构设计了一个频域特征解离模型,通过该模型可以分别解离出时域中叠加在原始OCT指纹图像上的信息(如区分性特征、无效特征和冗余特征).然后,我们让它学习不同的频域编码,并结合OCT指纹在时域中的重构编码形成相应的潜层编码.利用潜层编码,我们设计了一种用于区分表征攻击指纹和真实指纹的预测模型,用于表征攻击检测.在目前常用的OCT指纹数据集上的实验结果表明,我们的方法可以通过在频域中分离出一些叠加在时域中的无用干扰信息,从而有效地消除干扰.在实例方面,该方法的最小误差(Err.)为0.67%,与已有的基于时域的最优方法相比,最小误差降低了3.03%,性能提高了81.89%.

In automated fingerprint recognition systems(AFRSs),the development of fingerprint anti-spoofing ability is very crucial.Traditional fingerprints are usually obtained by surface fingerprint imaging,and such texture features are easy to be stolen.Fake fingerprints made of low-cost materials,such as artificial fingerprints made of 2D printing,silicone and other materials can easily attack these AFRSs.Therefore,using these traditional fingerprints for recognition will be difficult to detect presentation attacks.Existing research generally focuses on fingerprint modes with anti-counterfeiting features,such as high-resolution fingerprints with sweat gland characteristics and fingerprints with finger vein characteristics to develop presentation attack detection algorithms.This paper proposes a novel Optical Coherence Technology(OCT)-based fingerprint Presentation Attack Detection(PAD)method from the frequency domain to further improve the capability of fingerprint attack detection.OCT fingerprint imaging is a three-dimensional imaging technique that can capture subsurface fingerprint information beneath the fingertip's epidermis.An OCT fingerprint data is presented in the form of multiple cross-sectional images(i.e.B-scan),which can reflect multiple layers of biometric structure.It is very different from the surface image of a fingerprint.However,the existing PAD methods based on OCT fingerprint are traditional manual feature extraction methods and time-domain learning-based methods.Handcrafted extraction of fixed features in OCT fingerprint images is easily affected by noise,and these methods are not robust enough.Learning-based methods can learn the distribution of genuine and fake fingerprints and obtain more robust information representation in PAD.However,the information distribution in the image is superimposed,which may be ignored in the time-domain methods.Different from previous approaches,we first design a Frequency Feature Disentangling(FFD)model using convolutional neural networks and residual structures to decompose OCT-based fingerprint B-scans into four different frequency subbands like Discrete Wavelet Transform(DWT).Through such disentangling,information superimposed in the original image in the spatial domain(e.g.,discriminative PAD feature,invalid and redundant feature)can be separated respectively.We then let it learn different frequency codes to form their corresponding latent codes.Finally,the spoofness score which is used to distinguish PAs from bonafides is designed based on the latent codes.The experimental results on the commonly used OCT fingerprint dataset,evaluated on the dataset with 93200 bonafide B-scans from 137 fingers and 48400 B-scans from 121 PAs,show that our method can effectively preserve the most significant discriminative features and remove some useless interference information superimposed in the spatial domain by disentangling into the frequency domain for eliminating interference and effective PAD.In the performance comparison experiment with existing PAD methods,the proposed method achieves a minimum error(Err.)of 0.67%,which reduces the minimum error by 3.03%and improves the performance by 81.89%compared with the existing time-domain based state-of-the-art(SOTA)method,and there is a difference of only 0.4s in computing consumption.Additionally,we also compare the performance of the proposed method with the SOTA method in different attack materials.The proposed method achieves superior performance in both 2D and 3D attack materials,with a 3.72%reduction in Err.compared to the SOTA method specifically for 2D attack materials.