1 Introduction

The ECG is an inexpensive biological signal that reflects heart electrical activity by being monitored on the skin. In the last 30 years, ECG signal processing has developed into a vital part of clinical diagnosis and a major area of study. Individuals can be identified using their biometrics, which might include physical and behavioral traits [1]. Physical characteristics rarely or never undergo significant modifications. Features that are based on observable behaviors, however, may change over time and between contexts [2]. Physiological features like the face, fingerprint, hand geometry, iris, and retina, as well as behavioral features like the gait, signature, and speech, have all been used in the development of biometric recognition systems over the past 3 decades. The ECG signal [3] has recently been shown to be a useful biometric modality in clinical settings. Individual parameters including age, gender, location, size, and the heart’s structure contribute to the unique characteristics of an electrocardiogram. This is why biometric recognition systems have started utilizing ECG signals [4]. The requirement that the signals be taken from a living organism is the primary benefit of ECG-based biometric identification systems in terms of security [5].

ECG is the most developed and widely used method for analyzing physiological signals. The accuracy of ECG biometrics has been demonstrated in recent studies to be suitable for a variety of uses [6]. However, there are a number of challenges that ECG-based biometric systems must overcome. These include high intra-class variability due to heart rate and activities, a lack of studies that demonstrate the ability to discriminate in large scale datasets, and long-term changes in the features of an individual’s ECG [7]. Particularly susceptible to a drastic performance drop in uncontrolled acquisition circumstances are ECG biometric devices. In this research, a new deep learning-based method for biometric recognition based on electrocardiograms is proposed. Convolutional Neural Networks (CNNs) [8] is used to extract characteristics that allow for closed-set identification, identity verification, and periodic re-authentication. Deep-ECG can be performed with just one lead, or numerous leads can be fused for greater precision [9].

In addition, the ECG is a universal signal that indicates the presence of a living being and is extremely difficult to fake. Furthermore, there is currently no technology available that can simulate or generate a human ECG signal [10]. Therefore, ECG-based biometric recognition systems may find use in areas like security at check-out counters, airports, and borders. Several biometric recognition methods and applications based on ECG have been presented throughout the past 2 decades. Increases in processing speed and the advent of graphical processing units (GPUs) have led to the widespread adoption of deep learning-based algorithms for feature extraction and classification in biometric recognition [11]. In recent years, deep learning techniques like CNN models have been applied to ECG data for the purpose of biometric detection [12]. The ECG classification process is shown in Fig. 1.

Fig. 1
figure 1

ECG feature extraction and classification

Full size image

Every person’s heartbeat is different, and ECG authentication systems use sensors to capture that unique electrical activity. Usually worn as a patch on the skin, the sensors record the electrical cardiac signal and digitize it. The next step is to locate a match by comparing this template to a database of previously identified ECG templates. The individual is verified if there is a match. The heart is continually producing an ECG signal; therefore, authentication systems based on this data can enable continuous authentication and a high degree of accuracy. In addition to being a safe and easy way to confirm someone’s identity, ECG readings do not require any intrusive procedures. To make ECG-based authentication systems more efficient and accurate, deep learning models can be applied.

A deep learning algorithm can learn to identify individual-specific patterns in ECG signals by training on a massive dataset of ECG signals. Because of this, the model can correctly identify the subject whose ECG signal is being monitored and differentiate between other persons. There are a number of possible benefits of using ECG signals for identification rather than fingerprint recognition. To begin, ECG signals are specific to each person and can give very accurate results when used for authentication. Second, fingerprint recognition necessitates physical touch with a scanner, but ECG signals can be obtained non-invasively utilizing skin sensors. Third, unlike fingerprint recognition, which necessitates periodic re-scanning, ECG signals can offer continuous authentication, enabling continuous monitoring of an individual’s identity. The use of ECG data in authentication systems has the potential to be both quick and reliable in confirming a person’s identity.

A CNN is fed with a set of ECG signals together with the labels that go with them. Beginning with the input layer and concluding with the output layer, the CNN processes the ECG data through a succession of layers. Multiple units in each layer are linked to each other and to the layers below and above them via weights. After running the ECG signals through its paces, the CNN figures out how to tell authenticated signals apart from non-authenticated ones. The features are acquired by the process of convolution, which entails calculating the dot product between the input data and a small weighted matrix called a kernel at each point as it slides over the input data. The CNN adjusts the weights of the units as it processes the ECG data to reduce the discrepancy between the actual and predicted labels. Until the CNN learns to correctly categorize the ECG signals, it will keep doing this procedure, which is called backpropagation.

After training, the CNN can be used to identify verified ECG signals and categorize them. This is accomplished by feeding the updated ECG signal into the CNN and then making a prediction based on the CNN's output. A max-pooling layer, a fully connected layer, and a soft-max layer follow each of the five convolutional layers. The fully connected layer is uniformly designed across all networks. Rectified Linear Unit (ReLU) activation is provided to all fully connected layers since local response standardization increases computation time and memory use rather than improving the efficiency of our electrocardiogram collecting.

Images are captured, features are extracted from the images via processing, values are compared to determine an individual’s identity, and finally the two sets of values are compared. Single-modal and multi-modal biometric systems exist [13]. A single-modal biometric system is one that relies on just one biometric identifier. Despite the relative maturity of existing biometric technologies like face recognition, fingerprint recognition, and iris recognition [14], single biometric systems like fingerprint recognition have been severely impacted by the COVID-19 pandemic [15]. Facial recognition fails when the user is wearing a mask, while fingerprint detection is useless when the user is wearing gloves. The use of fingerprint and facial recognition technology as standalone solutions is becoming more and more problematic. Biometric identification is evolving from single to multi-modal as a result of the ongoing improvement and advancement of biometric identification technologies [16]. Using a well-thought-out architecture and fusion algorithm, it is possible to combine many biometric identifiers into a single scan, yielding complementing data and higher identification accuracy [17]. By integrating the strengths of various biometric identifiers and employing a variety of feature fusion methods, multi-modal biometrics can improve the reliability and safety of the identification process [18]. High identification accuracy, greater security, and a broader range of applications are only some of the benefits of multi-modal biometric identification technologies over single biometric identification methods. In this research, an Associated Priority-based Weighted Multi-Feature Vector model using Convolution Neural Network (APbWMFV-CNN) is proposed for ECG signal-based authentication.

2 Literature Survey

The ECG signals can be analyzed and valuable information can be extracted using deep learning algorithms. Applying a deep learning model to a massive ECG signal dataset is the usual procedure. After receiving fresh ECG signals, the model can use these to generate predictions or categorizations. Following this, a supervised learning method is used to train the deep learning model on this dataset. The algorithm tweaks the model’s internal parameters to minimize the prediction error on the training data.

There has been a recent uptick in research into biometric systems that use electrocardiograms. The two primary advantages of an electrocardiogram (ECG)-based biometric system are the ease of signal acquisition and the system’s resilience against counterfeit. The use of this biometric technology in personalized healthcare has the potential to automate authentication and subject identification. In this study, Jyotishi et al. [1] created a novel attention-based Hierarchical Long Short-Term Memory (HLSTM) model to learn the biometric representation that corresponds to an individual. An HLSTM model that can learn the ECG signal’s temporal fluctuations at many abstraction levels was proposed in this study. Users no longer have to worry about being dependent on LSTM networks in the long run; this fixes the issue. An individual’s most pertinent biometric data is educated into the model’s attention mechanism so that it can concentrate on ECG complexes. To obtain a more precise biometric representation, these ECG complexes are given greater weight. It is easier and more effective to use the proposed method since it does not depend on detecting fiducial points.

Many believe that biometric identification could solve the problem of insecure applications on the Internet of Things (IoT). Wu et al. [2] proposed an innovative ECG-based biometric identification method that could enhance the security of IoT patient monitoring systems. Since identification in biometric systems is both more difficult and more extensively researched than verification, the author devoted most of her attention to it. Users can prevent the cross-matching problem and secure their privacy by creating unique and irreversible templates for each enrollee using the principle of subspace oversampling. Without additional data for template building, subspace matching enables the identification of unknown people just from their beat bundles. To further ensure the database’s security, the proposed approach includes a mechanism to exclude subjects who have not registered, preventing the accidental association of an unidentified subject with an existing user.

Authentication plays a crucial role in ensuring the security of data transmitted between sensor nodes in a WBSN. Nodes in WBSNs that are capable of sensing physiological signals, including electrocardiograms (ECGs), continuously gather this information and use it for intrinsic liveness identification. When it comes to genuineness, these traits are perfect. There has been a lot of study on ECG-based intra-node authentication for WBSNs, but much less on protecting the ECG data, even though it is quite sensitive. This article proposes a private ECG-based authentication system based on a noninvertible transformation approach called manipulatable Haar transform (MHT) by Yang et al. [3]. With the proposed authentication mechanism in place, WBSNs can securely authenticate within nodes, protecting the privacy of patients’ vital health and identification data kept in electrocardiogram (ECG) files.

The use of multimodality in biometric authentication helps to circumvent the drawbacks of single-modality methods. An Android-based biometric identification system that uses both facial and vocal recognition was developed and implemented by Zhang et al. [4]. Considering the limitations of smart terminal hardware, such as random access memory (RAM), central processing unit (CPU), graphics processor unit (GPU), etc., which are unable to efficiently store and process large amounts of data, a face detection method is implemented to efficiently remove the redundant background of an image and minimize unnecessary information. With the introduction of a novel local binary pattern (LBP) coding method, the retrieved face feature is made even more trustworthy. By making improvements to the vocal activity detection (VAD) technique, which is an upgrade to the standard endpoint detection technology, the effectiveness of voice matching is improved since more information about voice mute and transitions is efficiently identified. To enhance the efficacy and precision of verification, the author presented an adaptive fusion method that harmoniously merges the advantages of facial and vocal biometrics. Public database cross-validation experiments demonstrate positive authentication performances when compared to other state-of-the-art methods.

The widespread adoption of simple facial authentication model raises privacy concerns due to the usage of personal biometric data as credentials. A privacy-preserving face authentication architecture should provide two key aspects, revocability and reusability, to safeguard users’ facial features. In the event of an authentication server compromise, revocability necessitates a method to remove or replace user credentials; reusability necessitates that the same credentials appear separately to non-cooperating applications. The complexity of the human face makes it difficult to simultaneously achieve these two qualities. Here, Lei et al. [5] introduced PrivFace, a quick, private face authentication method that works with both temporary and permanent biometric credentials. The main improvement is a new secure inner product protocol that measures face data similarity quickly using a lightweight random masked technique instead of laborious public-key cryptography procedures. The author provided a thorough analysis of the system’s security and demonstrated that the server is unable to learn private biological information about the user during the authentication process.

By learning a regularized mapping instead of a classification boundary, AuthNet—a novel framework for generic biometric authentication—introduced by Ali et al. [6]—performance and resilience are improved. We use simple and well-behaved probability to map the biometric traits of authorized and unauthorized individuals onto a latent space. Consequently, the author is free to employ simple and adaptable decision boundaries that are easy to alter. The author proved that compared to deep learning and conventional template-based authentication systems, regularizing the latent space to simple target distributions improves performance in terms of EER, accuracy, FAR, and GAR.

Although biometric methods are rapidly superseding passwords and tokens, they are still extensively utilized for authentication. Security measures for the biometric template have, thus, been the subject of intense research. This study by Tran et al. [7] presented a biometric authentication system that is powered by a lightweight AI and employs a binary representation of a biometric instance. To train a binary classifier, we will employ the binary strings that reflect biometric subjects’ intra-class and inter-class. The author accomplished this using two classifiers: a Support Vector Machine for fingerprint authentication and a Multi-layer Perceptron Neural Network for iris authentication. The authenticated biometric text is subsequently used to generate a hash value, which is then employed in a Zero-Knowledge-Proof Protocol to ensure confidentiality. The author devised a simple yet successful technique to improve the discriminativeness of the binary strings to enhance the classifier’s recognition; this technique is referred to as the Composite Features Retrieval.

Remote sensing, smart cities, and telehealth are just a few examples of how the IoT revolution is changing the world. People use IoT devices for everything from running their businesses to keeping tabs on their health. Massive volumes of useful data assets are produced by IoT devices. One area where sensor data may find use is in biometrics. Numerous attacks, including replication, repeated passwords, etc., can be launched against conventional biometric systems like PINs and passwords. Most current authentication strategies use motion-based sensors to compile individual profiles of users. The suggested method makes use of biological and motion sensors to provide robust multi-factor authentication. Batool et al. [8] presented an analytics framework for data from IoT sensors that can be used to build authentication models for users. To extract useful characteristics from the data, the author used a technique based on fiducial points. These characteristics serve as individual profiles for verification.

Kim et al. [8] introduced a novel LSTM Deep RNN architecture that relies on ECG classification. They then used various databases to evaluate the algorithm’s performance. The results show that the suggested technique is superior to and more successful than the conventional methods. The results show that the proposed model achieves better classification performance and accuracy (99.8%), recall (100%), and precision (0.99) than the conventional LSTM method. An encrypted biometric authentication system that uses electrocardiogram (ECG) signals and deep learning methods is presented by Prakash et al. [9] and is called BAED. By combining electrocardiography (ECG) impulses with deep learning techniques, BAED is able to identify individuals through biometric means. One of the main advantages of BAED is that it uses ECG signals, which are specific to each person and provide a great degree of security. Individuals can be accurately identified using ECG data using BAED’s deep learning techniques. The use of secure protocols to protect ECG data from unauthorized access is one of BAED’s advantages.

An individual’s privacy and the security of the system are both ensured by the system. In their proposal, Hosseinzadeh et al. [10] outlined a system to track the well-being of older persons by analyzing their biological and behavioral indicators. Using IoT devices like sensors and wearables, this system gathers data on a plethora of indicators, including heart rate, blood pressure, and sleep habits. It has the potential to monitor a person’s vitals in real time, enabling prompt medical attention in the event of an emergency. Biological and behavioral markers are just two of numerous signs that, when used together, paint a whole picture of a person’s health. Particularly helpful for elderly persons who live alone or have mobility issues, this technology also has the ability to remotely monitor them. Caregivers and family members can find some relief in this, as it reduces the likelihood of falls and other accidents.

3 Proposed Model

Pre-processing, feature extraction, matching, and decision-making are the four main modules of conventional biometric systems [19]. The system can be drastically changed by the technique used to extract features [20]. In this research, a deep learning model for feature extraction and selection is proposed for accurate extraction of ECG features and selecting the best feature set. After obtaining ECG signals, the signals are analyzed first. A dual-channel (CNN) was used to extract features from ECG signals. Before moving on to the fully linked layer, features are fused together and assigned weights based on the confidence each feature inspires. In this research, a framework is presented for fusing features from CNNs at the feature layer. The feature extraction, feature fusion, and classification recognition components make up the bulk of the framework.

3.1 ECG Data Processing

AI has become an essential tool for researching and retrieving various types of biological data. Practical deep learning algorithms have enabled the rapid development of biological and biomedical signal processing techniques in various fields, such as intelligence image analysis, surface-enhanced spectroscopic methods, and EEG. Biomedical data are often quite noisy and complex, making signal processing an essential step in evaluating and correctly restoring them. Building a complex strategy for peak performance sometimes necessitates a number of iterations, multiple computations, and computational equations for optimal choices. There is a potential alternative that makes use of conventional signal processing methods. DL makes an exact replica of the desired outcome, allowing for a quick and efficient service. Modern advances in learning strategies allow for the utilization of supervised, self-supervised, and unsupervised methods to accomplish the necessary four outcomes. Biomedical and biological sciences are seeing the growing importance of AI with the developments in biosensing, medical imaging, and informatics.

The literature reveals a great deal of heterogeneity in terms of data gathering hardware and methodologies. The amount of leads used, the duration of acquisitions, and how intrusive were all assessed. In addition, after reviewing the literature, we have come to the following conclusions about the best solutions: (1) off-the-person acquisitions, which bring ECG biometrics one step closer to practical, unrestricted applications; (2) one-lead setups; and (3) short-term acquisitions, which require fewer contact points, improve user acceptance of the data, and allow faster acquisitions, leading to a biometric system that is easy to use.

The most common type of neural network used for ECG processing at the moment is the feedforward neural network [21]. Using CNN to process ECGs complies with the principle of signal processing because it not only successfully reduces the dimension of huge data volume to small data amount, but it also effectively retains the properties of images [22]. CNN networks consist of an input layer, a hidden layer, and an output layer, just like any other neural network. The convolution layer is the backbone of a convolution neural network [23]; it is made up of numerous convolution units, and back propagation methods are used to fine-tune the parameters of each unit [24]. The primary purpose of the convolution operations performed by a multi-layer network is to extract features from ECG signals. The first few layers of a multi-layer network are best at extracting simple features [25]. The ReLU function of the activation function operation is the major focus of linear rectification, which allows the network’s non-linear mapping function to be realized, boosts the network’s expressive power, and has a positive impact on feature extraction. The proposed model framework is depicted in Fig. 2.

Fig. 2
figure 2

Proposed model framework

Full size image

3.2 Feature Engineering

Input to the pooling layer comes from multiple feature mapping and a pooling input process. Dimensions of features after convolution continue to partition the feature matrix into individual blocks to find its average or maximum; this may aid in dimensionality reduction, calculation speed for the network, and prevention of overfitting. The model’s inability to generalize is a result of overfitting the training data, which reduces its ability to handle data that differs from the training data. The pooling layer takes in a number of feature mappings and uses them as input. It is possible to minimize network processing, avoid overfitting, and lower dimensionality by dividing the feature matrix into numerous individual blocks and averaging or maximizing the image’s various dimensional properties. By integrating all local data and the feature matrix of each channel into vector representations, the fully connected layer computes the score for each final class.

Wearable gadgets have the potential to revolutionize authentication in the future by continuously capturing signals. We anticipate that these devices will soon dominate the sensing data market. In most cases, wearable devices are the primary means of obtaining the ECG signal, which then requires pre-processing to eliminate background noise. The ECG signal had to be separated into segments; without using any information about the position of the heartbeat, it was applied blindly into three-second segments. The goal is to create a method that can segment the data without wasting computational resources on locating the QRS complex. Only during feature extraction will the QRS complex be discovered.

Data augmentation is a different way to expand the amount of potential segments in this scenario. Following the segmentation procedure, feature extraction is employed. The features that emerge from this extraction are then processed to create a template that can be compared to the authorized user template. In particular, it uses the ECG to derive the user’s attributes for authentication purposes; nevertheless, the classification can be thrown off by outliers brought about by noise or acquisition-stage misplacement. A step to remove outliers is necessary in this regard. The last step is to use classification to differentiate between real and fake vital signal data.

Feature extraction is the most crucial part of the authentication procedure because it is at this stage that the user’s features are retrieved from the essential signal. The symbols used in the proposed model are indicated in Table 1.

Table 1 Proposed model nomenclature
Full size table

Initially, the ECG denoised data are considered as input and the data processing is applied. Data pre-processing involves in cleaning the data by applying standard deviations and mean calculation models that cleans the data. The data pre-processing is applied on the ECG dataset using Eqs. (1, 2, 3, and 4).

$$\text{mean}\left(r\right)=\frac{\sum_{i=1}^{M}s+(\text{attr}\left(i\right)+\text{attr}(i+1))}{L}$$
(1)
$$\text{Std}\left(r\right)=\sum_{i=1}^{M}\sqrt{\frac{{\sum }_{i=1}^{M}{(\text{getattr}\left(i\right)-\text{mean}\left(i\right))}^{2}}{L}}$$
(2)
$$\text{Dset}\left[N\right]=\sum_{r=1}^{N}\text{getrecord}\left(r\right)+\text{mean}\left(r,r+1\right)+\text{std}(r)$$
(3)
$$\text{DFset}\left[N\right]=\sum_{i=1}^{M}\text{max}(\text{Dset}\left(r\right))-\delta (r)\left\{\delta \left(r\right)\leftarrow \text{mean}\left(r,r+1\right) \text{iflen}(\delta \left(r\right)>0)\}\right.$$
(4)

where L is the total attributes in a record, i am the current attribute, r is the current record. δ is the special character in the dataset that will be removed with mean values.

Extracting features from ECG involves several processes, one of which is feature extraction. One example of a detected feature is the QRS complex in an ECG signal. The QRS complex, ST segment, and PR segment are all parts of a single heartbeat in an ECG. The RR interval, the amplitude of the P, R, and T waves, and other similar segments and intervals between fiducial points make up the features of an ECG signal. The feature extraction process will extract all the features from the ECG signal input and the process is performed using Eqs. (5) and (6)

$${\text{FextrSet}}\left[ N \right] = \mathop \sum \limits_{i = 1}^{M} \frac{{{\text{getattr}}\left( {i,i + 1} \right)}}{\lambda } + \max \left( {attr\left( i \right)} \right)\forall {\text{attr}}\left( i \right) \in {\text{Dset}}\left[ N \right]$$
(5)
$$\text{FeatSet}\left[N\right]=\prod_{i=1}^{M}\text{getVal}\left(\text{FextrSet}\left(i\right)\right)+\frac{\text{maxVal}\left(\text{FextrSet}\left(i\right)\right)-min(attr\left(i\right))}{\text{len}(\text{DFset})}$$
(6)

where λ is the total attributes considered from the dataset.

After extraction of feature set, each feature is allocated with weights. Weights represent the relativeness of the features to perform model training. The weight allocation is done to all the features in the extracted set. The process of weight allocation is performed using Eq. (7).

$$\text{WallocSet}\left[N\right]=\prod_{i=1}^{M}\underset{i\to N}{\text{lim}}{\left(\text{maxattr}(\text{FeatSet}\left(\text{i}\right))+\frac{\lambda }{i}\right)}^{2}$$
(7)
$$\text{Wset}\left[N\right]=\sum_{i=1}^{N}\text{mincorr}\left(\text{WallocSet}\left(i,i+1\right)\right)+\frac{\text{max}\left(\text{WallocSet}\left(i,i+1\right)\right)}{\lambda }$$
$${\text{mincorr}}(i,i + 1)\left\{ {_{{0\,\,otherwise}}^{{w \leftarrow flag + + if(corr\left( {{\text{WalloSet}}} \right)\; < Th)}} } \right.$$
(8)

where weights are allocated using a flag variable and features whose weight is assigned as 0, cannot be considered for processing. Corr() is the model for calculating the correlation factor among every two features in the feature extraction set.

The weighted feature set is considered and associated priority allocation is performed to the weakly correlated features. The weakly correlated features are the independent features that are most suitable for ECG feature vector generation. The process of associated priority allocation is performed using Eq. (9).

$${\text{PriorSet}}\left[ N \right] = \sum\limits_{{i = 1}}^{N} {\frac{{{\text{max}}({\text{Wset}}\left( {i,i + 1} \right))}}{\lambda }} + {\text{max}}({\text{corr}}({\text{Wset}}\left( {i,i + 1} \right)))\left\{ {_{{NormOtherwise}}^{{set\,priority \leftarrow Max\;if\;(Wset \ge Th)}} } \right.$$
(9)

Convolution layers, pooling layers, fully connected layers, and normalization layers are the typical components of a CNN’s hidden layers. The terms convolution and pooling functions are being used in place of the normal activation functions. A CNN hidden layer sits between its input and output layers; its artificial neurons in this layer process their weighted inputs and generate an activation function for its output. Two-thirds the size of the input layer plus the size of the output layer should be the target number of hidden neurons.

The convolution layer comes after a layer of M × M neurons. The output of the convolution layer will be of size (Mn + 1)(Mn + 1) if we employ a n × n filter τ. The pre-non-linear input to some unit Rγpq in current processing layer is calculated by adding the feature weights of the cells in the preceding layer: The process is performed using Eq. (10).

$${R}_{ij}^{\gamma }=\sum_{p=1}^{n-1}\sum_{q=1}^{n-1}{\tau }_{pq}{\mu }_{\left(i+p\right)(j+q)}^{\gamma -1}$$
(10)

The nonlinearity model is applied on the convolution using Eq. (11).

$${\mu }_{ij}^{\gamma }=\sigma ({R}_{ij}^{\gamma })$$
(11)

The kernel size is fixed using Eq. (12).

$$\text{Ksize}\left(\text{PriorSet}[N]\right)=\prod_{i=1}^{N}\text{max}(\text{PriorSet}\left(i\right))+\frac{\tau *(i+j)}{\mu }$$
(12)

The hidden layer processing is performed using Eq. (13)

$${\text{HidLyr}}[MXM] = \sum\limits_{{i = 1}}^{N} {R_{{ij}}^{\gamma } } *\frac{\mu }{\gamma } + \left( {{\text{max}}({\text{Ksize}}(i)} \right)) + \frac{{\tau *(p + q)}}{{{\text{len}}({\text{Ksize}})}}$$
(13)

where τ is the filter, (Mn + 1)(Mn + 1) is the CNN layers size, p,q and i,j are the features considered, µ is the neuron set, γ is the weighted feature component.

In mathematical information, feature vectors stand in for explanatory variable vectors. The feature vector set is generated from the associated priority set after hidden layer processing, considering the nest feature set for ECG-based biometric model. The feature vector set is generated using Eq. (14).

$$\text{FVTset}\left[M \right]=\sum_{i=1}^{N}\frac{\text{maxattr}(\text{HidLyr}\left(i\right))}{\tau }+\text{maxPriorSet}(\underset{i\to N}{\text{lim}}{\left({\mu }_{ij}^{\gamma }+\frac{\text{max}(\text{Wset}\left(i\right))}{{\tau }_{pq}{*\mu }_{pq}^{\gamma -1}}\right)}^{n}$$
(14)

4 Results

A fingerprint, electrocardiogram (ECG), iris, facial structure, or voice pattern are all examples of biometric data that can be used in a biometric system. Biometrics have many advantages over more conventional authentication systems, including being impossible to duplicate, share, forget, counterfeit, or alter. The use of biometrics has expanded beyond that of law enforcement in recent years. In addition, biometrics are being adopted by a growing number of enterprises as a means of controlling physical and digital access. Noise in sensed data, intra-class variants, inter-class similarities, non-universality, and spoof assaults are only some of the issues that might arise when using biometric modalities that only work with one another in real-world applications. A new ECG-based biometric system using relevant features that helps in selection of accurate features for better authentication rate is proposed.

Using a person’s unique physical behavioral traits, biometric identification can reliably and securely determine who that person is. To keep up with the ever-evolving threats posed by security breaches, biometric identification technology has undergone fast evolution over the past decade in search of more effective and convenient means of human authentication. The use of multiple biometric identifiers to verify a person’s identity has become increasingly popular as a result of developments in technology for sensors and signal processing methods. Signals from an individual’s ECG are a relatively new and exciting physiological input for use in highly secure biometric identification systems. Fiducial points (P, QRS, T, U, and V waves) can be found in every cardiac cycle, and these points serve as a representation of the electrical activity connected with the operation of the heart muscles.

Investigating ECGs as a non-invasive biometric, analogous to fingerprints, has become feasible due to developments in sensing technology. This bodes well for the widespread use of ECGs as biometrics. New long-term monitoring tools, including small wireless ECG body sensors, are being developed to supplement traditional off-the-person approaches. The electrical potential difference between electrodes placed near the heart can be measured by these sensors using a single lead. There are a number of variables, including the type of electrodes and the quantity and placement of leads, that cause wearable sensors to generate noisier signals than medical devices like Holter monitors. Wearable ECG systems, in contrast to medical ECG recorders, often employ one to three dry electrodes, with the first lead being utilized exclusively in mobile devices because of how straightforward it is to construct. Due to their more involved setup, longer and more detailed recording periods, and typically more trustworthy data, medical ECG recorders tend to outperform wearable devices.

After collecting ECG data, the following step in biometric processing is to extract features from the data. Based on inter-subject variability, these traits enable the detection of a particular individual based on certain parameters of the ECG. An essential part of pattern recognition is feature extraction. The success of feature selection method depends on the reliable identification of specific reference points, like the P wave, QRS complex, and T wave. When applied to these locations, these methods can extract biometric information in the form of interval, amplitude, angle, and area measurements. A lot of time-consuming feature engineering is required by these methods.

Several features can be derived from ECG signals and used as an input feature vector to a classifier to differentiate the identification of individuals after the raw ECG signals have been preprocessed using appropriate signal processing and compression techniques. The proposed model selects the most appropriate features used for accurate biometric model. The proposed model is compared with the traditional HLSTM model [1], Privacy-Preserving ECG-Based Authentication System for Securing Wireless Body Sensor Networks (PPECGbAS-WBSN) [2], and Android-based Multimodal Biometric Authentication System (AbMBAS) [4].

5 Dataset Description

Among the 47 participants evaluated by the BIH Arrhythmia Laboratory are 48 half-hour samples of two-channel ambulatory ECG recordings that are part of the MIT-BIH Arrhythmia Database. The study used a total of 4000 24-h ambulatory ECG recordings from patients. Outpatients made up about 40% of the population, while inpatients accounted for 60%. Twenty-three recordings were chosen at random from this set. The other 25 recordings were chosen from the same set to include less common but clinically significant arrhythmias that would not be adequately represented in a small random sample.

Using an 11-bit resolution over a 10-millivolt range, the recordings were digitalized at 360 samples/second per channel. The database contains computer-readable reference annotations for each beat, which were independently annotated by two or more cardiologists. Disputes were resolved to collect these annotations, which total around 110,000. The ECG denoised data are considered and the signal data are processed for performing feature extraction. The ECG data processing helps in gathering of clear signals for accurate ECG biometric detection. The ECG data processing time levels in milliseconds of the proposed and existing models are shown in Table 2 and Fig. 3.

Table 2 ECG data processing time levels in milliseconds
Full size table
Fig. 3
figure 3

ECG data processing time levels in milliseconds

Full size image

Data redundancy can be minimized with the aid of feature extraction. Finally, ECG data reduction aids model construction with less computational overhead, and boosts the rate of both learning and generalization. The text must be cleaned before its features may be extracted and used in a modeling context. The Table 3 and Fig. 4 show the feature extraction time levels in milliseconds of proposed and traditional models.

Table 3 Feature extraction time levels in milliseconds
Full size table
Fig. 4
figure 4

Feature extraction time levels in milliseconds

Full size image

The extracted features undergo weight allocation. Each features, based on the correlation attribute, weights are allocated to which priority can be assigned further. The weights represent the correlation weight that represents the dependant and independent features. The feature weight allocation accuracy levels in percentage of the proposed and existing models are represented in Table 4 and Fig. 5.

Table 4 Feature weight allocation accuracy levels in percentage
Full size table
Fig. 5
figure 5

Feature weight allocation accuracy levels in percentage

Full size image

After weight allocation to the features, associated priority allocation is performed to assign a priority to a feature in the ECG signal set. The ECG biometric model works purely on considering the associated priority feature set. The associated priority allocation accuracy levels in percentage of the proposed and existing models are shown in Table 5 and Fig. 6.

Table 5 Associated priority allocation accuracy levels in percentage
Full size table
Fig. 6
figure 6

Associated priority allocation accuracy levels in percentage

Full size image

An artificial neural network’s hidden layer is in between its input and output layers; its artificial neurons in this layer process their weighted inputs and generate an activation function for its output. The complexity and high performance of neural networks can be traced back to their hidden layers. They can simultaneously carry out tasks like data transformation and the generation of automatic features. The hidden layer processing accuracy levels in percentage for the existing and proposed models are shown in Table 6 and Fig. 7.

Table 6 Hidden layer processing accuracy levels in percentage
Full size table
Fig. 7
figure 7

Hidden layer processing accuracy levels in percentage

Full size image

Since a CNN’s output is a feature vector, feeding it a vector set as input would result in a feature vector representing the ECG signal data. To calculate the feature vector, the similarity measurement is verified and the final feature vector set is generated that is used for designing the accurate ECG-based biometric model. The Table 7 and Fig. 8 show the feature vector generation time levels in milliseconds of the existing and proposed models.

Table 7 Feature vector generation time levels in milliseconds
Full size table
Fig. 8
figure 8

Feature vector generation time levels in milliseconds

Full size image

6 Conclusion

In today’s data-driven world, people of all backgrounds are interested in the best methods for verifying identities and keeping sensitive data safe. While biometric identification is now the most practical and safe method of establishing one’s identity, this method is limited in its ability to accommodate the increasingly diverse and complicated authentication circumstances that are becoming the norm. This research offers a CNN-based biometric technique that extracts and selects the features from ECG for accurate detection of humans To extract information from ECG data, CNN analyzes collections of QRS complexes. This research presents a feature layer fusion approach using a CNN, bringing together the best feature set. To maximize the useful feature information, this technique uses a residual structure and incorporates a self-attention mechanism that assigns a weight to each feature as it is updated by the feature fusion module. After data collection, feature extraction and feature reduction are the most crucial steps. The most distinguishing characteristic values are chosen. It would be reduced to a negligible quantity of data if these traits were sufficiently large. The proposed model performance can be measured by applying the generated feature set to a variety of closed-set identification, identity verification, and periodic re-authentication tasks while employing large sample sets collected under realistic situations. In this research, an (APbWMFV-CNN) is proposed for ECG signal-based authentication. The proposed model selects the most appropriate features used for accurate biometric model. The proposed model achieves 98.4% accuracy in generating the weighted multi feature vector. The proposed model feature subset reduces the feature set with best feature set. In future, optimization techniques can be applied on the feature set to perform feature dimensionality reduction with reduced feature set.