A novel two-dimensional ECG feature extraction and classification algorithm based on convolution neural network for human authentication
A novel two-dimensional ECG feature extraction and classification algorithm based on convolution neural network for human authentication
复制标题
一种基于卷积神经网络的新型二维心电图特征提取和分类算法，用于人体认证
DOI：
10.1016/j.future.2019.06.008
复制
发表时间：
2019-12-01
影响因子：
7.5
通讯作者：
Wang, Kuanquan
中科院分区：
文献类型：
--
作者：
Hammad, Mohamed;Zhang, Shanzhuo;Wang, Kuanquan
基于生物特征的认证系统通常是比传统系统更好的安全解决方案，传统系统严重依赖密码，个人身份证号码或智能卡。心电图（ECG）是近年来最有前途的基于生物特征的身份认证方法之一，因为与其他生物特征不同，它可以确保被认证人的存活性。在本文中，我们提出了一种新的身份验证系统，使用一种有效的特征检测算法和卷积神经网络（CNN）的基础上的ECG人体认证。我们的系统通过两个主要阶段处理ECG信号：特征检测阶段和认证阶段。在特征检测阶段，首先进行预处理，以尽可能多地去除噪声并拉直ECG信号。然后，提出的扫描和去除方法被用来提取的主要特征，具有负峰值，高品位的噪声和基线漂移的信号与更高的精度比现有的算法。在认证阶段，我们提出了一个12层CNN来认证ECG信号。我们还引入了一个新的数据库（MWM-HIT数据库），它适用于训练和验证认证系统。此外，我们使用了PTB、CYBHi和MIT-BIH心律失常数据库中的所有记录，以比较拟议系统和其他系统。对于从收集的数据库中检测所有记录上的所有峰，我们分别实现了97.92%、96.96%和98.79%的平均准确度、灵敏度和阳性预测性，并且对于10秒的导联II ECG信号，检测MIT-BIH数据库上的所有峰的准确度为99.96%、灵敏度为99.99%和阳性预测性为99.98%。该模型是能够认证的ECG信号的等误差率（EER）为1.63%，4.47%，和4.86%时，使用铅II从PTB，CYBHi和收集的数据库，分别。所提出的系统是高度可用的实时认证系统。(C)2019 Elsevier B. V.版权所有。
A biometrics-based authentication system is usually a better security solution than traditional systems which are heavily reliant on passwords, personal identification numbers or smart cards. Electrocardiogram (ECG) is one of the most promising approaches for biometrics-based authentication in recent years, because, unlike other biometrics, it assures aliveness of the person being authenticated. In this paper, we present a novel authentication system using an efficient feature detection algorithm and a convolutional neural network (CNN) based on ECG for human authentication. Our system processes ECG signals through two main phases: a feature detection phase and an authentication phase. In the feature detection phase, preprocessing was performed first to remove as much noise as possible and straighten the ECG signals. Then, the proposed scanning and removing methods are used to extract the main features from the signals that have negative peaks, high-grade noise and baseline drifts with higher accuracy than existing algorithms. In the authentication phase, we proposed a 12-layer CNN to authenticate the ECG signals. We also introduced a new database (MWM-HIT database), which is suitable for training and validating authentication systems. In addition, we used all records from the PTB, CYBHi, and MIT-BIH arrhythmia databases for comparison between the proposed system and other systems. We achieved an average accuracy, sensitivity and positive predictivity of 97.92%, 96.96% and 98.79%, respectively for detecting all peaks on all records from the collected database and an accuracy of 99.96%, sensitivity of 99.99% and positive predictivity of 99.98% for detecting all peaks on the MIT-BIH database for ten seconds of lead II ECG signals. The proposed model is able to authenticate the ECG signals with an equal error rate (EER) of 1.63%, 4.47%, and 4.86% when using lead II from PTB, CYBHi and the collected databases, respectively. The proposed system is highly usable in a real-time authentication system. (C) 2019 Elsevier B.V. All rights reserved.