ArticlesAll Issue
ArticlesEnd-to-End Data Authentication Deep Learning Model for Securing IoT Configurations
• Mohamed Hammad1, Abdullah M. Iliyasu2,3,4,*, Ibrahim A. Elgendy5,6, and Ahmed A. Abd El-Latif7,*

Human-centric Computing and Information Sciences volume 12, Article number: 04 (2022)
https://doi.org/10.22967/HCIS.2022.12.004

Abstract

Compared to other biometrics, electrocardiograms (ECGs) have gained widespread acceptability as mediums for validating animateness in numerous security applications, especially in new and emerging technologies. Our study utilizes this important trait to advance available machine and deep learning ECG authentication systems by leveraging the use of edge computing servers that offer connection to Internet of Things (IoT) devices while maintaining access to computational and storage resources. Specifically, in our proposed technique, the preprocessing, feature extraction and classification routines are combined into one unit, while individual ECG signals from the database are directly fed into a convolutional neural network (CNN) model and subsequently classified as an accepted or unaccepted (i.e., rejected) class. Additionally, we tailor our authentication system as a cost-efficient one focused on reducing latency, which makes it ideal for applications on edge computing platforms. To validate our proposed model, we applied it on standard ECG signals from the Physikalisch-TechnischeBundesanstalt (PTB) database where outcomes of 99.50%, 99.73%, 100%, and 99.78%, respectively that are reported for accuracy, precision, recall, and F1-score indicate the tenability of deploying our technique in real-time authentication systems. Furthermore, we present discussions regarding the performance of the model relative to recent techniques that are built on traditional machine and deep learning techniques.

Keywords

Deep Learning, Biometric Authentication, ECG, Information Security, IoT, Edge Computing, Industrial Data Integration

Introduction

Recently, increases in incidents of counterfeiting have forced a scale down in the acceptability of traditional authentication methods such as security tokens, passwords, personal identification numbers (PINs), etc. Instead, the use of biometric authentication is now prevalent. Even so, traditional biometric markers such as fingerprint, face recognition, iris recognition and ear recognition are susceptible to influences of wide-ranging factors such as universality, uniqueness, performance, measurability, acceptability, circumvention, etc. [1]. In addition to providing good compromise in these factors and unlike the traditional biometric modalities, such as fingerprints, which do not always imply consciousness, the use of electrocardiogram (ECG) signals as biometric markersprovidesa good test of animateness.
The ECG measures the rhythm associated with the pumping of blood in and out of the heart [2]. Recently, available ECG sensors have become smaller in size, which makes it easier to integrate them into wearable devices and connected vehicles [3]. Additionally, minimal computation and template size are among other reasons attributed to the popularity and utility of ECGs in different fields [4,5]. In [6], these properties are exploited for the secure identification of drivers and passengers in vehicle access control.
Meanwhile, the ECG is considered pivotal in the diagnosis of various cardiac conditions, including arrhythmias, and other coronary disorders. A typical cardiac cycle of an ECG signal comprises a P-wave, a QRS-complex, and a T-wave [2]. In addition, there are three essential segments or intervals known as ST-segment or ST-interval, PR-segment or PR-interval, and QT-interval. The atrial depolarization caused by atrial contraction is traced as P-wave while ventricular depolarization or contraction is indicated by the QRS-complex segment. Afterward, both ventricles come back to relaxation states, which are known as T-wave [7]. The PR-interval, which is also known as the PR-segment, is the line between the start of the P-wave and the Q-wave, whereas the QT-interval is the line between the start point of the QRS-complex and that of the T-wave. Similarly, the ST-interval, which is also known as the ST-segment, is the line between S-wave and T-wave. Fig. 1 illustrates a typical ECG signal as well as its peaks and intervals. These characteristics have been utilized in wide-ranging applications, including improvement of security in body-area network sensors (BANs) such as smart wristbands and heart pacemakers [8].

Fig. 1. Illustration of all peaks and intervals on an ECG signal [20].

Lately, there has been an upsurge in the use of these signals in applications across different medical Internet of Things (IoT) platforms [9]. However, there is still a shortage of such applications using ECG for authentication [10]. Meanwhile, whereas cloud computing approaches are more widespread for handling IoT applications, they are known to suffer from high latency to applications, impact of considerable traffic overhead, etc., all of which are detrimental to real-time and delay-sensitive applications [11]. Unlike these past efforts, our study presents a deep learning-based ECG authentication approach for edge computing platforms where, unlike in IoT, cloud resources and services are moved to edge nodes [12].
Broadly speaking, previous studies on ECG-based authentication can be classified as either conventional machine learning techniques [1316] or deep learning-based approaches [1723]. In the former, the ECG signals pass through many steps (typically for preprocessing of the ECG data, extracting and selecting the ECG features and classification) before the authentication is accomplished. These steps complicate and overburden the targeted authentication process. Moreover, most of these approaches reported below-par authentication performance when working on other databases as well as when evaluated on larger datasets that overfitting had been reported [1316].
To overcome these limitations, we propose a deep learning model (DLM) approach to human authentication system using ECG signals. As one of its many upsides, DLM techniques subsume many of the earlier mentioned steps of traditional machine learning approaches. Deep learning is a class of machine learning where networks are capable of unsupervised learning from unstructured or unlabeled data. These networks facilitate the gradual extraction of higher-level features from raw input data. The most widespread used DLM method, known as convolutional neural network (CNN), has its origins and best performance in image analysis [24]. Recently, CNN-based DLMs have become popular across many fields including medicine [2528], computer vision [2932], remote sensing [33,34], and many other interesting applications [3538].
To encapsulate, whereas the use of DLMs in biometric authentication, particularly CNN [39] is not new, very few of these methods are based on ECG authentication [1723]. Even so, they are encumbered by the following shortcomings.

Temporal and resource overhead associated with the need for preprocessing, segmentation and, QRS detection stages (as in [1316]).

Sensitivity to quality of ECG signals (as in [1719]).

Need for a separate classifier for authentication, which complicates the already complex process [2023].

Working on more than one lead ECG, which imposes additional hardware susceptibility (as in [17,20]).

To ameliorate the enumerated deficiencies of both the machine and deep learning approaches, our proposed method presents the following contributions.

An efficient CNN-based DLM is proposed to automatically authenticate the ECG signals for efficient identity validation.

Since ECG data suffer from various noise artifacts due to unconstrained signaling environments and considering the challenges of collecting all in one environment, this study employs an end-to-end DLM to eliminate the need for a preprocessing stage.

An edge-computing framework is integrated into the proposed model as a cost-efficient remedy to lower latency wherein storage and processing capabilities are provided at the edge of the radio access network that is in close proximity to mobile users. Consequently, our strategy facilitates real-time and delay-sensitive applications for execution.

As noted earlier, there has been a recent surge in the integration of IoT and wireless technologies into a wide range of systems such as healthcare, pattern recognition, industry [3941]. These devices including sensors are infused in automatic car tracking adapters, Google Glass, blood glucose monitors, ECGs, continuous glucose monitors, oximeters, insulin pens, temperature monitors, etc. Correspondingly, numerous applications and multimedia services that require supplementary computations and high data communication are increasing rapidly. Nevertheless, these applications are still resource-limited with restricted computational power and energy. Edge computing, which is described as a network architecture that provides computing and data storage at edge server nodes [12], is considered a solution that can address the limitations of individual devices. Moreover, it is known to offer lower latency and high bandwidth [12], which supports its implementation in real-time applications. Consequently, computationally intensive services of IoT devices can be offloaded and executed on this server through a wireless channel. This will mitigate the load demand while also helping to extend battery life [12].
Our study advances previous machine and deep learning ECG authentication efforts by leveraging the use of edge computing servers that can be connected to IoT devices while maintaining access to computational and storage resources. Fig. 2 illustrates the integration of ECG authentication in a car control system which is considered a daily life application. First, the user’s ECG signal is acquired and sent to the edge computing server through mobile devices where the proposed model is trained, and relevant data are stored. Next, if the user is authenticated, the edge server will send an automated signal to the car sensor which transfers further control to the user. Otherwise, the edge server will reject the user’s request and subsequently alert appropriate authorities to prevent the vehicle from being stolen. Therefore, the overall performance of these systems is improved since computationally-intensive operations of the model are processed at the edge server. Table 1 presents a definition of parameters used in this example and subsequent sections of this study.

Fig. 2. Integration of ECG authentication in a car control system.

The rest of this paper is structured as follows. In Section 2, we present the outline and other details of our proposed CNN model whose validation via a range of performance evaluation experiments are presented in Section 3. These results are subsequently analyzed and discussed alongside related work in Sections 4 and 5, respectively.

Table 1. Definition of parameters
Parameter Definition
N Element number or the number of sets applied in the 5-fold
x Signal
f Filter in the input signal (x)
P Precision
R Recall
EER Equal error rate
ROC  Receiver operating characteristic
F1 F1-score
FP False positive
FN False negative
TP True positive
TN True negative
M Number of classes

Related Research

As highlighted in our Introduction, previous studies on ECG-based authentication can be broadly classified as either classical machine learning (CLM)-based techniques [1316,42] or DLM-based approaches [1723,4347]. This section provides an overview of recent studies in both categories.

Classical Machine Learning-based Techniques
Classical machine learning (CLM) refers to techniques that pass through all steps of the machine learning pipeline [48]. In this kind of system, the ECG signals pass through many steps before the authentication is accomplished. The first stage entails removal of noise and other artifacts while also enhancing the input ECG signals in order to make them suitable for further processing. This step is usually called the preprocessing step. The second step is focused on extracting and selecting the prominent features from the preprocessed ECG signal. These features should be unique for each signal and this step is called feature extraction step. Finally, the features are fed into the classifier in the classification step where the authentication is done.
Most CLM techniques adopt basic frameworks outlined above. For instance, in [13], the authors introduced a feature extraction and a continuous authentication method by proposing one-dimensional multi-resolution local binary patterns (1DMRLBP) and sequential sampling feature extraction for one-dimensional signals. This system updates decision thresholds and sample sizes during run-time. As reported in [13], it accounts for the quantization error needed to tolerate noise and extracts local and global signal morphology. For validation, the authors reported using data from 290 subjects from the Physikalisch-TechnischeBundesanstalt (PTB) database and recorded an equal error rate (EER) of 10.10% for ECG signal authentication. Similarly, Safi et al. [14], presented a feature extraction approach that they called pulse active ratio (PAR) for ECG-based authentication using ECG fiducial points as a feature vector. In this scheme, each feature vector from the test dataset is compared to all feature vectors in the training database with Euclidean distance set as the similarity measurement to generate matching scores. If the feature vectors from the test and training dataset are the same, the authentication is considered successful. They reported experiments involving 112 individuals from the PTB dataset of which they reported that 98 individuals exhibited arrhythmia beats while the remainder were healthy. They also recorded an EER of 9.89 % and 19.15% for normal and arrhythmia beats, respectively. In their contribution, Goshvarpourand Goshvarpour[42], developed an expert system for human authentication based on ECG signals. They executed their authentication protocol by combining the ECG characteristics with information theoretic (IT) considerations. In the last step, using 5-fold cross-validation, they applied k-nearest neighbor (kNN) to identify individuals and reported an average accuracy of 97.6%. In their contribution, Ivanciu et al. [49], presented an ECG-system based on Siamese neural networks for authentication. They used the Siamese network to simplify the training process. In addition, they deployed the system on private cloud to ensure its security and scalability. They reported a sensitivity of 87.3% and overall accuracy of 86.47%. In [50], the authors introduced an ECG authentication system based on incremental learning to identify the ECG signal of each subject under several conditions. They employed support vector machine (SVM) for authentication and reported the best accuracy of 99.4%. Similarly, in [51], the authors presented an ECG authentication system based on nonlinear normalization according to various physiological conditions. They reported an accuracy of 99.05% and 88.14% for resting and non-resting conditions, respectively.
To conclude, we also note that despite the performance recorded, authentication performance in most of the highlighted CLM studies, are below-par when working on other databases or when assessed on large datasets, which also leads to overfitting. Exploiting their potency, nowadays, DLMs have been proposed to overcome most of the problems faced by CLM methods across different domains and applications as highlighted in the sequel.

Security Scenarios with CNN-based DLMs
Deep learning is widely regarded as the advanced solution to most machine learning problems. As reviewed earlier, most DLM are built on CNNs, and they are credited with some of the stellar performance recorded in various fields including biometric authentication. For example, in [17], the authors used CNN to deploy an authentication algorithm based on more than one ECG lead. Their algorithm recorded an EER of 1.36% for authentication using the PTB database. In [18], the authors presented a CNN-based multi-modal authentication system based on ECG and fingerprints. Their technique used a pre-trained VGG-Net model and transfer learning technique by using the output of the last fully connected layer as features for the input ECG image. Following this, an improved version of bio-hashing strategy [16] was employed to protect the features. Finally, an external classifier called Q-Gaussian multiclass support vector machine (QG-MSVM) was used for authentication. They reported an EER of 3.20% for the PTB database.
Furthermore, in [19] the authors used CNN to generate ECG features that they protected using matrix operations [16]. In the end, they employed a separate classifier (QG-MSVM) for authentication and reported an EER of 3.5% using images in the PTB database. In another effort, Hammad et al. [20] combined machine learning features with deep learning features of ECG for human authentication using CNN. They employed scan and removal techniques to extract the ECG features and CNN for classification. They recorded an EER of 1.63% for ECG authentication. In their contribution, Chu et al. [23] used the parallel multi-scale one-dimensional residual network for ECG authentication. Specifically, in the preprocessing step, they located the R-peak and used it with the whole ECG signal to extract the feature. After that, they used a residual network with a new loss function to extract the features and train the network. They used all records in the PTB database and reported an EER of 0.59%.
Similarly, Zhang et al. [43] presented a wavelet domain multiresolution CNN (MCNN) for ECG identification. The first step of their technique used blind segmentation to process the data and process first blind segmentation and applied an auto-correlation operation on the transformed data to remove the difference resulting from the use of randomly chosen signal segments. They reported an average identification accuracy of 93.5%. Similar to that effort, Abdeldayem and Bourlai [44], presented a human identification system based on ECGs where the input signal was segmented to enrich its original informational content. Subsequently, they generated spectral correlation images that were fed to the CNN for identification. They reported an accuracy of 95.6% with false acceptance and rejection rates of 0.2% and 0.1%, respectively. Similarly, in [45], the authors presented a CNN protocol for human identification based on ECG. They employed two CNN models, one for feature extraction and another for identification a combination that recorded an average identification accuracy of 94.3%.
In their contribution, Abd El-Rahiem and Hammad [52] presented a multi-fusion ECG authentication system based on internal deep fusion algorithm. Transfer learning technique is used to extract the deep features of ECG signals in this system. After that, they used SVM as a separate classifier for authentication and reported final authentication accuracy of 99.4%. Finally, AlDuwaile and Islam [53] employed CNN and a single heartbeat to develop an ECG recognition system. They reported accuracy of 99.90%, 98.20%, and 94.18% for different databases. Unlike most of the enumerated DLMs, our proposed approach is built on an end-to-end structure that eliminates the need for different machine learning stages while recording better accuracy. Details of this proposed model are presented in the next section.

Materials and Methods

Structurally, our proposed CNN-based DLM (or simply CDM) combines the hitherto separate stages (i.e., preprocessing, feature extraction and classification) of conventional machine learning techniques into a compact unit. Like most of the studies highlighted in the previous section, ECG biometric data from the PTB database [54] was employed for all stages of this study. This dataset [54] comprises of ECG signals from 290 subjects, which are separately fed into our proposed CNN model. Finally, the proposed model classifies the acquired ECG signal as an accepted or unaccepted (i.e., rejected) class. For example, in our earlier demonstration of vehicle access control systems based on ECG authentication in Fig. 2, the system grants access depending on the identification of the driver, i.e., whether he/she is accepted as an authorized driver or not. Fig. 3 presents a general outline depicting the integration of the DLM in the ECG authentication framework

Fig. 3. Block diagram outlining implementation of the proposed model.

Dataset and Implementation Framework
As mentioned earlier, the PTB dataset [54] is used throughout this study. This is a widely used multi-session dataset made up of 549 records from 290 subjects whose ages range from 17 to 87 years of which 28% were female and 72% were male. Each subject has signal duration of between 1 to 5 records. Each recording consists of 15 measured signals: 12 regular leads and 3 Frank ECG leads with a sampling frequency of 1,000 Hz. The specifications of the PTB recorder are as follows:

±16 mV as an input voltage with 16 input channels and 100 Ω input resistance

16-bit resolution with 0 to 1 kHz bandwidth.

Maximum noise voltage of 10 μV (pp).

The noise level recording during signal collection. Furthermore, the diagnostic classes of the subjects in the dataset are summarized in Table 2 [54].

Additionally, we used two-second Lead II ECG signals made up of 2,000 sample pulses in total. Finally, we normalized each record using Z-score normalization with standard deviation and mean set at 1 and 0, respectively.
Results emanating from the implementation of our proposed model are compared alongside standard studies published in the literature as reported later in Section 4 and discussed in Section 5.

Table 2. PTB classes
Class Number of subjects
Myocardial infarction 148
Myocardial hypertrophy 7
Myocarditis 4
Heart failure 18
Dysrhythmia 14
Bundle branch block 15
Valvular heart disease 6
Miscellaneous 4
Healthy 52

Proposed CNN-based Deep Learning Model
CNN has been credited for the remarkable success of DLMs. Since its debut for applications in pattern recognition and computer vision, it has been deployed in numerous fields, including biometrics, authentication, etc. Typically, a CNN consists of different layers: convolutional, fully connected, and pooling. The convolutional layer is used to extract features from the input that are fed into the pooling layer where some parameters may be reduced. Finally, the fully connected layer converts the features matrix to a vector and combined, these features together to create a model.
Fig. 4 depicts details of the CNN architecture for our proposed model comprising of five convolution processes that consist of three maxpooling, three dropout, two fully connected layers as well as a softmax layer that is employed in the authentication stage, whose output is the final decision, i.e., accept or reject.

Fig. 4. CNN operation of the proposed model.

Probing further, we now note that in our model, each ECG record (with 2,000 samples) is fed to our network for training and testing. The proposed CNN model comprises of five 1D convolutional layers that are actuated using a rectified linear unit (ReLU) activation function [54] and three maxpooling layers that are followed by two fully connected layers and a softmax activation function, which is used to transform the outputs of the classifier into probabilities for each class of the two classes (i.e., accept or reject). In accomplishing this, the convolution operation is executed using (1).

$Y_p=\displaystyle\sum_{k=0}^{N-1}x_k f(n-k)$(1)

where N, x, and f are the element number, signal, and filter in the input signal (x), respectively and the vector y represents the output signal. We start with the 1D convolutional layers with 64 filters whose size increases sequentially to 128 filters in the last layers (to extract more features from the input). The filter size is set at 3 with a step-wise stride. We also pool the output of every convolutional layer output at a stride of 2. Further, a dropout with a 0.2 probability rate is used to prevent the network from overfitting [[55]] by ignoring some randomly selected neurons during the training. It is noteworthy that the dropout is important in large networks that have complex fitting but few labeled samples. In addition to serving as a classifier, the fully connected layer maps the dimensions, i.e., high to low, as presented in Table 3. For authentication, we employed the softmax function defined in (2) because it is differentiable, and it optimizes the cost function, and its outcome produces the final decision as presented in Fig. 3.

$σ(\overrightarrow Z)i=\frac{e^{Z_i}}{\displaystyle\sum_{j=1}^k e^{Z_j}}$(2)

where $Z_i$ is the input vector to the softmax function, k is the number of classes and $\displaystyle\sum_{j=1}^k e^{Z_j}$ is a normalization term.

Training
Stochastic gradient descent with momentum [56] is used for training our model with a batch size of 10 samples. Furthermore, hyper-parameters comprising of weight decay, learning rate and number of epochs were set at 0.00005, 0.0001 and 100, respectively. Subsequently, these parameters were adjusted iteratively to obtain optimum performance.

Table 3. Composition of different layers of proposed model
Layer number Layer name Size of filter Stride Number of kernels Size of output
1 Convolution  3 1 64 1998×64
2 Convolution 3 1 64 1996×64
3 Maxpooling 2 2 998×64
4 Dropout 　 Rate=0.2 　 998×64
5 Convolution 3 1 128 996×128
6 Convolution 3 1 128 994×128
7 Maxpooling 2 2 497×128
8 Dropout 　 Rate=0.2 　 497×128
9 Convolution 3 1 128 494×128
10 Maxpooling 2 2 　 247×128
11 Fully connected 512 256
12 Dropout 　 Rate=0.2 　 256
13 Fully connected 448 224
14 Softmax 　 　 2 2

Testing
Testing was performed after each run of the training epoch on our model. We divided the data into three parts: 60% for training, 30% for validation, and 10% for testing. Additionally, a 5-fold cross-validation approach [57] is employed for training and testing sets during authentication. In this approach, the ECG data are randomly shuffled and divided into five equally-sized subsets. After that, we ran five experiments where 90% of the data was reserved for training and validation and the remaining 10% for testing. Finally, we evaluated the final performance of the method by calculating the mean of all folds.
To ascertain the performance of our CNN model, we considered two factors: evaluating the impact of the proposed model using a 5-fold approach [57] and the authentication performance of the proposed method relative to established ECG authentication methods. Algorithm 1 outlines the execution of the above procedure while Algorithm 2 summarizes steps for the execution of the CNN model.

Algorithm 1. Execution of proposed protocol
1: Input: Raw input data x= x1, x2, …, xn; % with 2000 samples for each record
2: CNN ← x; % the input data are directly fed into the CNN network to retrieve the extracted feature vectors
3: F = F1, F2, …, Fn; % the extracted feature vectors are mapped into high-dimensional space
4: for each ECG signal starting from i = 1 to the length of ECG signal (2000 samples)
5: extract CNN features → F;
6: end for
7: Ftrain, Ftest, Ltrain, Ltest ← F, L; % split features and labels into train and test subsets
8: Output: F to SoftMax activation function; % transform the outputs of the classifier into probabilities for each output of the two classes (Accept or Reject)
Algorithm 2. Execution of proposed CNN model
Function CNN (x):
x ← set the input data.
Conv 1 ← RELU_Activation_Func (Conv1D)
Conv 2 ← RELU_Activation_Func (Conv1D)
Conv 2 ← Max_Pool1D (Conv 2)
Conv 2 ← Dropout (0.2)
Conv 3 ← RELU_ Activation_Func (Conv1D)
Conv 4 ← RELU_ Activation_Func (Conv1D)
Conv 4 ← Max_Pool1D (Conv 4)
Conv 4 ← Dropout (0.2)
Conv 5 ← RELU_ Activation_Func (Conv1D)
Conv 5 ← Max_Pool1D (Conv 5)
Fully connected ← Reshape (Conv 5)
Fully connected ← Dropout (0.2)
Fully connected ← RELU_ Activation_Func (Fully connected)
Output ← Softmax_Func (Fully connected)
Return Output
End Function

Experimental Validation

Our proposed model was executed on an Intel Core i7-6800K CPU with NVIDIA GTX 1080ti GPU and 16 GB RAM workstation equipped with MATLAB R2018a open-source Deep Learning toolbox. The training time of the CNN model was 0.5 hour while the prediction lasted 0.03 seconds.

Performance Metrics
For performance assessment, we employed standard metrics, such as accuracy (A), precision (P), recall (R), EER, receiver operating characteristic (ROC) curve and F1-score (F1) that depend on false positive (FP), false negative (FN), true positive (TP), and true negative (TN) rates as defined in the equation matrix in Fig. 5 [26] where N and M represent the number of sets applied in the 5-fold validation and the number of classes, respectively. We note that the ROC curve is a graph used to show the performance of a statistical model with two output classes (such as the Accept or Reject) outcome of our model. It provides a trade-off between specificity and sensitivity and classifiers with high accuracy tend to have curves closer to the left corner [[58]].

Fig. 5. Equation matrix for binary quality metrics. Adapted from [26].

Authentication Results
Using the performance metrics presented in Fig. 5, in this subsection an overview of the outcomes of the authentication process is presented and subsequently discussed in the next section. Table 4 presents the authentication results for the proposed system per fold for the PTB database by computing the accuracy, precision, recall and F1-score in each fold. Furthermore, these outcomes are presented as confusion matrices for each fold in Fig. 6 from where an average accuracy of 99.50% is reported. Based on this figure, 99.4% of the genuine classifications are correctly verified (i.e., true positive) while a meagre 0.04% of impostor cases are wrongly verified as genuine classifications, i.e., false positive classifications (or type II error).

Table 4. Authentication results per fold on PTB database
Number of folds Accuracy (%) Precision (%) Recall (%) F1-score (%)
1 99.3 99.6 99.98 99.7
2 98.7 98.9 99.98 99.4
3 100 100 100 100
4 99.5 99.7 99.99 99.8
5 100 100 100 100

Fig. 6. Confusion matrix for each fold of proposed model: (a) 1-Fold, (b) 2-Fold, (c) 3-Fold, (d) 4-Fold, and (e) 5-Fold.

Fig. 7. ROC curves of each fold of proposed model.

Additionally, Fig. 7 presents the ROC curves for the proposed CNN model in each fold, while the variation in the EER (computed as a percentage rate when FP=FR) for our proposed model during the 5-fold validation is presented in Fig. 8. Here, it is noteworthy that an average EER of 0.47% is reported for the proposed model, while, as presented in Fig. 9 (i.e., plots of accuracy and logarithmic loss), it can be deduced that the model converges after 50 training epochs. Consequently, this setting is adopted for the remaining experiments that are reported.

Fig. 8. Variation in EER (%) for different number of folds in proposed model.

Discussion of Results

Whereas analysis of the results presented earlier in Section 4 can provide deductions regarding the efficiency and robustness of the proposed authentication method, in this section we present an analysis of such outcomes. From the authentication results in Table 4 and the confusion matrices in Fig. 6, we can observe that the proposed method attained high accuracy in each fold (i.e., in some cases maximum accuracy of 100% was reported), which, as previously reviewed in Section 2, manifests the robustness of the proposed model and its capacity to overcome overfitting that is ascribed to CLM models,. Additionally, the ROC curves reported in Fig. 7 present evidence that the proposed model has high authentication in each fold as exhibited by the number of genuine acceptance rate relative to the number of false acceptance rate in each fold. Finally, the variations in EER reported in Fig. 8 support our claims regarding the accuracy of the proposed model since low EER is recorded in each fold.

Fig. 9. Accuracy and loss curves for the proposed CNN technique.

To undertake an effective performance analysis, the authentication performance obtained using the PTB database based on [1316] and deep learning techniques [1723] highlighted earlier in Section 2 (presented in Table 5), alongside outcomes from our proposed model. Therefrom, we note the better performance (in bold) of out technique. Moreover, for a balanced discussion, we start by highlighting the key features of the methods used in the assessment (i.e., as presented in Table 5).
Unlike the deep learning approaches mentioned in Table 5 [1720, 23], our proposed approach is built on an end-to-end structure that eliminates the need for different machine learning stages. Furthermore, our system is less complex than other systems because only one ECG lead was used. Additionally, unlike the enormous size of the filters in the convolutional layers (e.g., 5×1 and 8×1) employed by the model in [17], which increases the computational cost, our model utilizes a reasonable number of filters throughout the network (i.e., 3×1 filters). Moreover, unlike methods (such as [18, 20, 21]) that recorded low accuracy using noisy data, as an end-to-end structure, our model can manage such noisy signals. This eliminates the need for a preprocessing stage, and it works directly on the input ECG signal instead. Furthermore, unlike our proposed model, previous studies in [1720, 22, 23] reported the use of separate classifiers, which exacerbates computational complexity and cost.
As noted earlier in Section 3, the use of edge computing servers to execute and evaluate the proposed deep learning model, provides access to radio networks to enhance computational and storage capabilities. Therefore, the overall performance of these systems is improved since computationally-intensive operations of the model are processed on the edge server. In this regard, edge computing models are considered prominent solutions that can offer lower latency and high bandwidth to support their implementation in real-time applications.
We conclude our discussion by once again highlighting the contributions of our study vis-à-vis its merits as highlighted in the reported discussions in this section:

As demonstrated throughout the study, our model is less complex than other authentication approaches because it is conceived and built as a completely end-to-end structure.

This end-to-end structure supports direct application on the ECG signal, thus circumventing the need for any preprocessing stage. This facilitates dealing with noisy ECG signals.

Finally, compared to previous DLMs, our proposed model exhibits better robustness and effectiveness in overcoming traditional drawbacks, such as overfitting.

Table 5. Comparison of different authentication algorithms based on PTB database (see text for further details)
ECG authentication framework Study Methodology Performance (EER, %)
Machine learning methods Louis et al. [13] Local binary patterns 10.1
Sequential sampling and bagging
Safie et al. [14] Pulse active ratio 9.89 (for normal class)
Euclidean distance 19.15 (for abnormal class)
Gürkan et al. [15] Autocorrelation/Discrete cosine transform 2.84
Mel-frequency cepstrum coefficients
Linear discriminant analysis
Three-nearest-neighbor
Hammad et al. [16] Improved bio-hashing 32
Matrix operation 14
Feed forward neural network
Deep learning models Labati et al. [17] CNN 2.9
Hammad et al. [18] Two-dimensional CNN 3.2
Hammad and Wang [19] One-dimensional CNN 3.5
Hammad et al. [20] Scanning and removing techniques 1.63
CNN
Hammad et al. [22] CNN 1.53 (using CNN)
Residual CNN (ResNet) 1.39 (using ResNet)
Chu et al. [23] Parallel multi-scale one-dimensional residual network 0.59
Proposed 1D convolutional neural network 0.47

Concluding Remarks

Our study advances previous machine and deep learning technique for ECG-based human authentication by leveraging the use of edge computing servers that can be connected to IoT devices while maintaining access to computational and storage resources. The proposed model employs a CNN-based DLM to realize an end-to-end structure that eliminates the need for any handcrafted preprocessing, feature extraction and classification, procedures that are known to impact on the computational complexity of the system. Consequently, our model exemplifies the potentials of integrating the edge-computing frameworks to enhance cost effectiveness and efficiency in terms of lower authentication time. Using the PTB ECG database outcomes of experiments reported support claims that the proposed method significantly outperforms existing classical machine and deep learning models in terms of authentication EER. These outcomes indicate improved generalization and robustness for our proposed approach.
With appropriate modifications, some real-life applications of the proposed model, such as a car access control system, customer authentication for the banking system, and applications in many medical systems are interesting directions to implement the proposed study. Despite its potentials, however, some drawbacks of the proposed model include its lack of cancellability, i.e., the ability to protect the ECG feature templates [59, 60]. Overcoming these shortcomings of the proposed model are among our short-term goals whereas, in the future, we intend to use template protection techniques to safeguard the system against spoof attacks and explore other applications for the model across other biometrics systems. Furthermore, we will assess security threats and propose mechanisms against attacks as well as applications in advanced IoT infrastructures [46].

Acknowledgments

None.

Author’s Contributions

Conceptualization, MH, AMI. Funding acquisition, AMI. Investigation and methodology, MH, AMI, AAA. Project administration, AMI, AAA. Resources, AMI. Supervision, AMI, AAA. Writing of the original draft, MH, AMI, AAA. Writing of the review and editing, MH, IAE, AMI. Software, IAE. Validation, IAE, AAA. Formal analysis, AMI, IAE. Data curation, MH, IAE. Visualization, MH, AAA. All the authors have proofread the final version.

Funding

This study is sponsored by the Prince Sattam Bin Abdulaziz University, Saudi Arabia via the Deanship for Scientific Research funding for the Advanced Computational Intelligence & Intelligent Systems Engineering (ACIISE) Research Group (Project No. 2020/01/12173).

Competing Interests

The authors declare that they have no competing interests.

Author Biography

Affiliation: Information Technology Department, Faculty of Computers and Information, Menoufia University, 32511, Egypt
He received his MSc degree in 2015, Information Technology Department, Faculty of Computers and Information, Menoufia University, Egypt. He worked as a demonstrator and assistant lecturer in Faculty of Computers and Information, Menoufia University, Egypt since April 2012 till now. He is currently a PhD candidate at the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China. His research interests include Computer Vision, Machine Learning, Pattern Recognition and Biometrics.

Name: Ibrahim A. Elgendy
Affiliation: Computer Science Department, Faculty of Computers and Information, Menoufia University, 32511, Egypt
He received his M.Sc. degree from Computer Science Department, Faculty of Computers and Information, Menouﬁa University, Egypt, in 2016. He worked as a demonstrator and assistant lecturer in Faculty of Computers and Information, Menouﬁa University, Egypt since April 2012 till now. He is currently pursuing the Ph.D. degree from the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China. His research interests include Cloud Computing, Mobile Edge Computing and distributed computing.

Name: Abdullah M. Iliyasu
Affiliation: Electrical Engineering Department, Prince Sattam Bin Abdulaziz University, Al-Kharj 11942, Saudi
He received the M.E., Ph.D., and Dr.Eng. degrees in computational intelligence and intelligent systems engineering from the Tokyo Institute of Technology, Japan. He is the Principal Investigator and Team Leader of the Computational Intelligence and Intelligent Systems Research Group at the College of Engineering, Prince Sattam Bin Abdulaziz University in the Kingdom of Saudi Arabia. He has to his credit more than 100 publications traversing the areas of computational intelligence, quantum cybernetics, quantum image processing, quantum machine learning, cyber and information security, hybrid intelligent systems, health informatics, and electronics system’s reliability.

Name: Ahmed A. Abd El-Latif
Affiliation: Mathematics and Computer Science Department, Faculty of Science, Menoufia University, Shebin El-Koom 32511, Egypt
He (SMIEEE, MACM) received his B.Sc. with honor rank in Mathematics and Computer Science in 2005 and M.Sc. degree in Computer Science in 2010, all from Menoufia University, Egypt. He received his Ph. D. degree in Computer Science and Technology at Harbin Institute of Technology, China in 2013. He is currently an associate professor at Faculty of Science, Menoufia University, Egypt. His areas of interests include quantum computing, cybersecurity, and IoT security.

References

[1] A. S. Alghamdi, K. Polat, A. Alghoson, A. A. Alshdadi, and A. A. Abd El-Latif, “A novel blood pressure estimation method based on the classification of oscillometric waveforms using machine-learning methods,” Applied Acoustics, vol. 164, article no. 107279, 2020. https://doi.org/10.1016/j.apacoust.2020.107279
[2] K. A. Abuhasel, A. M. Iliyasu, and I. N. Alquaydheb, “Reappraising the Impact of environmental stresses on the useful life of electronic devices,” Journal of Advanced Computational Intelligence and Intelligent Informatics, vol. 20, no. 4, pp. 640-651, 2016. https://doi.org/10.20965/jaciii.2016.p0640
[3] A. S. Alghamdi, K. Polat, A. Alghoson, A. A. Alshdadi, and A. A. Abd El-Latif, “Gaussian process regression (GPR) based non-invasive continuous blood pressure prediction method from cuff oscillometric signals,” Applied Acoustics, vol. 164, article no. 107256, 2020. https://doi.org/10.1016/j.apacoust.2020.107256
[4] A. Savvas, “Halifax Bank trials heart rate technology to authenticate customers,” 2015 [Online]. Available: https://www.computerworld.com/article/3556854/halifax-bank-trials-heart-rate-technology-to-authenticate-customers.html.
[5] A. Alghamdi, M. Hammad, H. Ugail, A. Abdel-Raheem, K. Muhammad, H. S. Khalifa, and A. Ahmed, “Detection of myocardial infarction based on novel deep transfer learning methods for urban healthcare in smart cities,” Multimedia Tools and Applications, 2020. https://doi.org/10.1007/s11042-020-08769-x
[6] C. Burt, “EKG biometrics from B-Secure to be featured in 2020 car models.,” 2019 [Online]. Available: https://www.biometricupdate.com/201905/ekg-biometrics-from-b-secur-to-be-featured-in-2020-car-models.
[7] M. Hammad, A. Maher, K. Wang, F. Jiang, an M. Amrani, “Detection of abnormal heart conditions based on characteristics of ECG signals,” Measurement, vol. 125, pp. 634-644, 2018. https://doi.org/10.1016/j.measurement.2018.05.033
[8] C. L. Chen and C. T. Chuang, “A QRS detection and R point recognition method for wearable single-lead ECG devices,” Sensors, vol. 17, no. 9, article no. 1969, 2017. https://doi.org/10.3390/s17091969
[9] P. Huang, L. Guo, M. Li, and Y. Fang, “Practical privacy-preserving ECG-based authentication for IoT-based healthcare,” IEEE Internet of Things Journal, vol. 6, no. 5, pp. 9200-9210, 2019.
[10] C. L. P. Lim, W. L. Woo, S. S. Dlay, D. Wu, and B. Gao, “Deep multiview heartwave authentication,” IEEE Transactions on Industrial Informatics, vol. 15, no. 2, pp. 777-786, 2018.
[11] I. A. Elgendy, W. Z. Zhang, C. Y. Liu, and C. H. Hsu, “An efficient and secured framework for mobile cloud computing,” IEEE Transactions on Cloud Computing, vol. 9, no. 1, pp. 79-87, 2018.
[12] M. A. Jan, W. Zhang, M. Usman, Z. Tan, F. Khan, and E. Luo, “SmartEdge: an end-to-end encryption framework for an edge-enabled smart city application,” Journal of Network and Computer Applications, vol. 137, pp. 1-10, 2019.
[13] W. Louis, M. Komeili, and D. Hatzinakos, “Continuous authentication using one-dimensional multi-resolution local binary patterns (1DMRLBP) in ECG biometrics,” IEEE Transactions on Information Forensics and Security, vol. 11, no. 12, pp. 2818-2832, 2016.
[14] S. I. Safie, J. J. Soraghan, and L. Petropoulakis, “Electrocardiogram (ECG) biometric authentication using pulse active ratio (PAR),” IEEE Transactions on Information Forensics and Security, vol. 6, no. 4, pp. 1315-1322, 2011.
[15] H. Gurkan, U. Guz, and B. S. Yarman, “A novel biometric authentication approach using electrocardiogram signals,” in Proceedings of 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Osaka, Japan, 2013, pp. 4259-4262.
[16] M. Hammad, G. Luo, and K. Wang, “Cancelable biometric authentication system based on ECG,” Multimedia Tools and Applications, vol. 78, no. 2, pp. 1857-1887, 2019.
[17] R. D. Labati, E. Munoz, V. Piuri, R. Sassi, and F. Scotti, “Deep-ECG: convolutional neural networks for ECG biometric recognition,” Pattern Recognition Letters, vol. 126, pp. 78-85, 2019.
[18] M. Hammad, Y. Liu, and K. Wang, “Multimodal biometric authentication systems using convolution neural network based on different level fusion of ECG and fingerprint,” IEEE Access, vol. 7, pp. 26527-26542, 2018.
[19] M. Hammad and K. Wang, “Parallel score fusion of ECG and fingerprint for human authentication based on convolution neural network,” Computers & Security, vol. 81, pp. 107-122, 2019.
[20] M. Hammad, S. Zhang, and K. Wang, “A novel two-dimensional ECG feature extraction and classification algorithm based on convolution neural network for human authentication,” Future Generation Computer Systems, vol. 101, pp. 180-196, 2019.
[21] M. G. Kim and S. B. Pan, “Deep learning based on 1-D ensemble networks using ECG for real-time user recognition,” IEEE Transactions on Industrial Informatics, vol. 15, no. 10, pp. 5656-5663, 2019.
[22] M. Hammad, P. Pławiak, K. Wang, and U. R. Acharya, “ResNet: attention model for human authentication using ECG signals,” Expert Systems, vol. 38, no. 6, article no. e12547, 2021. https://doi.org/10.1111/exsy.12547
[23] Y. Chu, H. Shen, and K. Huang, “ECG authentication method based on parallel multi-scale one-dimensional residual network with center and margin loss,” IEEE Access, vol. 7, pp. 51598-51607, 2019.
[24] A. Sedik, M. Hammad, F. E. Abd El-Samie, B. B. Gupta, and A. A. Abd El-Latif, “Efficient deep learning approach for augmented detection of coronavirus disease,” Neural Computing and Applications, 2021. https://doi.org/10.1007/s00521-020-05410-8
[25] M. Amrani, M. Hammad, F. Jiang, K. Wang, and A. Amrani, “Very deep feature extraction and fusion for arrhythmias detection,” Neural Computing and Applications, vol. 30, no. 7, pp. 2047-2057, 2018.
[26] A. Sedik, A. M. Iliyasu, A. El-Rahiem, M. E. Abdel Samea, A. Abdel-Raheem, M. Hammad, J. Peng, F. E. Abd El-Samie, and A. Ahmed, “Deploying machine and deep learning models for efficient data-augmented detection of COVID-19 infections,” Viruses, vol. 12, no. 7, article no. 769, 2020. https://doi.org/10.3390/v12070769
[27] S. L. Oh, V. Jahmunah, C. P. Ooi, R. S. Tan, E. J. Ciaccio, T. Yamakawa, M. Tanabe, M Kobayashi, and U. R. Acharya, “Classification of heart sound signals using a novel deep WaveNet model,” Computer Methods and Programs in Biomedicine, vol. 196, article no. 105604, 2020. https://doi.org/10.1016/j.cmpb.2020.105604
[28] T. Ozturk, M. Talo, E. A. Yildirim, U. B. Baloglu, O. Yildirim, and U. R. Acharya, “Automated detection of COVID-19 cases using deep neural networks with X-ray images,” Computers in Biology and Medicine, vol. 121, article no. 103792, 2020. https://doi.org/10.1016/j.compbiomed.2020.103792
[29] A. Fourcade and R. H. Khonsari, “Deep learning in medical image analysis: a third eye for doctors,” Journal of Stomatology, Oral and Maxillofacial Surgery, vol. 120, no. 4, pp. 279-288, 2019.
[30] N. Xie, X. Li, K. Li, Y. Yang, and H. T. Shen, “Statistical karyotype analysis using CNN and geometric optimization” IEEE Access, vol. 7, pp. 179445-179453, 2019.
[31] R. Singh, R. Khurana, A. K. S. Kushwaha, and R. Srivastava, “Combining CNN streams of dynamic image and depth data for action recognition,” Multimedia Systems, vol. 126, pp. 313-322, 2020.
[32] F. Zhu, X. Kong, H. Fu, and Q. Tian, “A novel two-stream saliency image fusion CNN architecture for person re-identification,” Multimedia Systems, vol. 24, no. 5, pp. 569-582, 2018.
[33] Y. You, C. Lu, W. Wang, C. K. Tang, “Relative CNN-RNN: learning relative atmospheric visibility from images,” IEEE Transactions on Image Processing, vol. 28, no. 1, pp. 45-55, 2018.
[34] L. Zhong, L. Hu, H. Zhou, and X. Tao, “Deep learning based winter wheat mapping using statistical data as ground references in Kansas and northern Texas, US,” Remote Sensing of Environment, vol. 233, article no. 111411, 2019. https://doi.org/10.1016/j.rse.2019.111411
[35] W. Nie, K. Wang, Q. Liang, and R. He, “Panorama based on multi-channel-attention CNN for 3D model recognition,” Multimedia Systems, vol. 25, no. 6, pp. 655-662, 2019.
[36] S. K. Tripathy and R. Srivastava, “A real-time two-input stream multi-column multi-stage convolution neural network (TIS-MCMS-CNN) for efficient crowd congestion-level analysis,” Multimedia Systems, vol. 26, no. 5, pp. 585-605, 2020.
[37] A. Yadav and D. K. Vishwakarma, “A deep learning architecture of RA-DLNet for visual sentiment analysis,” Multimedia Systems, vol. 26, pp. 431-451, 2020.
[38] J. Zhao, X. Mao, and L. Chen, “Speech emotion recognition using deep 1D & 2D CNN LSTM networks,” Biomedical Signal Processing and Control, vol. 47, pp. 312-323, 2019.
[39] H. Zhang, A. Jolfaei, and M. Alazab, “A face emotion recognition method using convolutional neural network and image edge computing,” IEEE Access, vol. 7, pp. 159081-159089, 2019.
[40] L. Li, K. Ota, and M. Dong, “Deep learning for smart industry: efficient manufacture inspection system with fog computing,” IEEE Transactions on Industrial Informatics, vol. 14, no. 10, pp. 4665-4673, 2018.
[41] P. Li, Z. Chen, L. T. Yang, Q. Zhang, and M. J. Deen, “Deep convolutional computation model for feature learning on big data in Internet of Things,” IEEE Transactions on Industrial Informatics, vol. 14, no. 2, pp. 790-798, 2017.
[42] A. Goshvarpour and A. Goshvarpour, “Human identification using information theory-based indices of ECG characteristic points,” Expert Systems with Applications, vol. 127, pp. 25-34, 2019.
[43] Q. Zhang, D. Zhou, X. Zeng, “HeartID: a multiresolution convolutional neural network for ECG-based biometric human identification in smart health applications,” IEEE Access, vol. 5, pp. 11805-11816, 2017.
[44] S. S. Abdeldayem and T. Bourlai, “A novel approach for ECG-based human identification using spectral correlation and deep learning,” IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 2, no. 1, pp. 1-14, 2019.
[45] Y. Li, Y. Pang, K. Wang, and X. Li, “Toward improving ECG biometric identification using cascaded convolutional neural networks,” Neurocomputing, vol. 391, pp. 83-95, 2020.
[46] A. Ahmed, R. Latif, S. Latif, H. Abbas, and F. A. Khan, “Malicious insiders attack in IoT based multi-cloud e-healthcare environment: a systematic literature review,” Multimedia Tools and Applications, vol. 77, no. 17, pp. 21947-21965, 2018.
[47] C. Tong, B. Liang, Q. Su, M. Yu, J. Hu, A. K. Bashir, and Z. Zheng, “Pulmonary nodule classification based on heterogeneous features learning,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 2, pp. 574-581, 2021.
[48] M. Wierzbinski, P. Plawiak, M. Hammad, and U. R. Acharya, “Development of accurate classification of heavenly bodies using novel machine learning techniques,” Soft Computing, vol. 25, no. 10, pp. 7213-7228, 2021.
[49] L. Ivanciu, I. A. Ivanciu, P. Farago, M. Roman, and S. Hintea, “An ECG-based authentication system using Siamese neural networks,” Journal of Medical and Biological Engineering, vol. 41, no. 4, pp. 558-570, 2021.
[50] J. Kim, G. Yang, J. Kim, S. Lee, K. K. Kim, and C. Park, “Efficiently updating ECG-based biometric authentication based on incremental learning,” Sensors, vol. 21, no. 5, article no. 1568, 2021. https://doi.org/10.3390/s21051568
[51] H. B. Hwang, H. Kwon, B. Chung, J. Lee, and I. Y. Kim, “ECG authentication based on non-linear normalization under various physiological conditions,” Sensors, vol. 21, no. 21, article no. 6966, 2021. https://doi.org/10.3390/s21216966
[52] A. El-Rahiem and M. Hammad, “A multi-fusion IoT authentication system based on internal deep fusion of ECG signals,” in Security and Privacy Preserving for IoT and 5G Networks. Cham, Switzerland: Springer, 2022, pp. 53-79.
[53] D. A. AlDuwaile and M. S. Islam, “Using convolutional neural network and a single heartbeat for ECG biometric recognition,” Entropy, vol. 23, no. 6, article no. 733, 2021. https://doi.org/10.3390/e23060733
[54] R. Bousseljot, D. Kreiseler, and A. Schnabel, “Nutzung der EKG-signaldatenbank CARDIODAT der PTB über das Internet,” Biomedical Engineering/Biomedizinische Technik, vol. 40, no. s1, pp. 317-318, 1995. https://doi.org/10.1515/bmte.1995.40.s1.317
[55] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929-1958, 2014.
[56] I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the importance of initialization and momentum in deep learning,” in Proceedings of the 30th International Conference on Machine Learning (ICML), Atlanta, GA, 2013, pp. 1139-1147.
[57] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2nd ed. New York, NY: John Wiley and Sons, 2001, pp. 55-88.
[58] M. H. Zweig and G. Campbell, “Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine,” Clinical Chemistry, vol. 39, no. 4, pp. 561-577, 1993.
[59] A. D. Algarni, G. M. El Banby, N. F. Soliman, F. E. A. El-Samie, and A. M. Iliyasu, “Efficient implementation of homomorphic and fuzzy transforms in Random-Projection encryption frameworks for cancellable face recognition,” Electronics, vol. 9, no. 6, article no. 1046, 2020. https://doi.org/10.3390/electronics9061046
[60] B. A. El-Rahiem, M. Amin, A. Sedik, F. E. Samie, and A. M. Iliyasu, “An efficient multi-biometric cancellable biometric scheme based on deep fusion and deep dream,” Journal of Ambient Intelligence and Humanized Computing, 2021. https://doi.org/10.1007/s12652-021-03513-1

Mohamed Hammad1, Abdullah M. Iliyasu2,3,4,*, Ibrahim A. Elgendy5,6, and Ahmed A. Abd El-Latif7,*, End-to-End Data Authentication Deep Learning Model for Securing IoT Configurations, Article number: 12:04 (2022) Cite this article 2 Accesses