Multichannel EfficientNet B7 with attention mechanism using multimodal biometric- based authentication for ATM transaction

Abstract

Automatic Teller Machine (ATM) offers rapid and user-friendly avenues to reach their bank accounts and engage in financial operations. A crucial component of ATM security is the “Personal Identification Number (PIN) or password”. This PIN or password serves as a fundamental element in safeguarding and preserving customers’ financial data from unauthorized entry. Within the financial realm, an ongoing necessity exists to enhance security measures. In the realm of identity verification, modern ATM systems traditionally require the combination of an access card and the input of a PIN. However, the landscape has evolved with the emergence of cutting-edge biometric authentication methods like fingerprint scanning, retina recognition, and facial identification. These innovations have significantly mitigated the security vulnerabilities previously associated with ATMs. To surmount such challenging factors, a novel multimodal biometric-based authentication is introduced for ATM transactions. Traditionally, the MultiBank Provider (Pvt Company) provides an ATM card with all bank access for an individual. With the help of ATM machines, multimodal authentication is accomplished by using the Multichannel EfficientNet B7 with Attention Mechanism (MEB7-AM), in which each channel carries information about each image from the Face, Retina, Fingerprint, and spectrogram. Once it is done, a single pin is required to select the bank. Further, from the selected bank with proper credentials, the money is withdrawn from the ATM machine. Lastly, the efficacy of the model is analyzed using various measures and compared among existing methodologies. Therefore, the proposed system provides the precise results of better authentication for ATM machines.

Keywords

Authentication for ATM transaction biometric information multichannel EfficientNet B7 with attention mechanism single card multiple bank

1. Introduction

In the past, authentication methods primarily relied on two-factor authentication, employing either passwords or PIN codes alongside smart cards [1]. This approach, combining knowledge and ownership factors, is deemed more secure than single-factor authentication [2]. The knowledge factor pertains to user-specific information like passwords and Personal Identification Number (PIN) codes, while the ownership factor involves possession of items such as smart or Automatic Teller Machine (ATM) cards [3]. An example of two-factor authentication is ATM transactions, which necessitate both knowledge of a PIN code or password and possession of a specific ATM card [4]. Despite the convenience of online banking, security and privacy concerns have deterred some customers from adopting these services. The surge in technology-enabled banking operations has captured the interest of criminal elements [5]. Two-Factor Authentication (2FA) commonly entails a dual-layer verification process: the user acquires a code via email, SMS, or a token and subsequently inputs this code alongside their username, and occasionally a PIN or password, to gain access to the service [6]. Consequently, when choosing an authentication mechanism for an online banking system, it’s essential to account for elements like security, cost-efficiency, compatibility, and user-friendliness [7]. However, the vulnerability of information stored within smart cards became evident, leading to the reconsideration of authentication schemes using smart cards due to password guessing, impersonation, and related attacks [8]. Conversely, asymmetric strategies encompass tasks such as exponentiation, point multiplication, and pairing, which become unfeasible in situations with limited resources. In contrast, authentication systems based on symmetric keys, leveraging shared keys [9], offer a more fitting approach for resource-constrained environments, owing to their streamlined nature and reliance on less computation-intensive operations like symmetric encryption, hash functions, and XOR. Nonetheless, such schemes often lack defense against certain attacks or encounter correctness-related issues [10]. Passwords or PINs enable access to private information in ATMs, but remain susceptible to various attacks [11]. As an alternative, biometrics have gained traction. Biometrics rely on human physiological or behavioral features for authentication, offering heightened information security. This makes biometrics a popular alternative to traditional password or PIN-based authentication systems [12].

Recognizing the limitations of password-centric authentication methods involving smart cards, a range of three-factor biometrics-driven authentication strategies have been adopted [13]. These three-factor authentication approaches hinge on the principles of knowledge, ownership, and inherence factors [14]. Inherence factors utilize biometric features such as fingerprints or iris offering greater security than passwords due to their difficulty to forge, share, or break [15]. Individuals are identified through what they possess (ID cards, smart cards), what they know (PINs, passwords), and unique biometric traits. While security systems based on possession can be lost, stolen, forged, or duplicated [16], systems based on what individuals know are vulnerable to being forgotten, shared, stolen, guessed, or hacked. On the other hand, systems based on unique traits are less easily compromised. Multimodal biometrics enhance spoof attacks, security against vulnerabilities, inter-class, and intra-class variations, and non-universality [17].

The key contributions of the proposed ATM transaction security framework, which relies on the utilization of Multichannel EfficientNet B7 in conjunction with Attention Mechanism via deep networks, are outlined as follows.

To develop an ATM transactions security framework using Multi-modal biometric authentication for enhancing the security of ATM transactions. This approach combines multiple biometric factors (face, retina, fingerprint, and spectrogram) to ensure a higher level of security and accuracy, reducing the risks associated with traditional card and PIN methods.

To propose MEB7-AM is designed to efficiently process and analyze the biometric information from various sources. This mechanism ensures that each channel’s data contributes to the authentication process, thereby increasing accuracy and reducing false positives.

To done the successful multimodal authentication, the system streamlines the transaction process. Users only need to enter a single PIN to select the desired bank. Once the bank is selected and proper credentials are provided, the ATM machine facilitates the withdrawal of money seamlessly.

To assess the system’s efficacy using a comprehensive set of performance measures. By comparing the results with existing authentication methodologies, the system’s accuracy, speed, and reliability can be quantified, demonstrating its superiority and effectiveness. To assess the system’s efficacy using a comprehensive set of performance measures. The system’s accuracy, speed, and reliability can be quantified by comparing the results with existing authentication methodologies to demonstrate its superiority and effectiveness.

The subsequent sections expound upon the security framework for ATM transactions, employing Multi-modal biometric authentication through deep networks. The overview of conventional research findings is encapsulated in Section 2, along with a concise discussion of its merits and limitations. Section 3 delves into security concerns related to banking transactions, particularly the use of a single ATM card for multiple banks. Moving to Section 4, the model’s design is outlined, encompassing the collection of biometric-based data for authentication in ATM transactions. The utilization of multichannel EfficientNet B7 with an attention mechanism for ATM transaction authentication is detailed in Section 5. The outcomes of experiments, including comparisons with various established systems and performance metrics, are presented in Section 6. The final segment, Section 7, concludes by highlighting that the proposed framework yields precise user recognition performance.

2. Literature review

In 2018, Ogbanufe and Kim [18] proposed a research model, incorporating hypotheses that were subsequently evaluated to compare how individuals perceive different payment authentication methods, their level of trust in online stores, and their willingness to continue using website accounts linked to these methods. An experimental approach was adopted to empirically test these hypotheses. The findings indicated that biometric authentication significantly impacted individuals’ perceived utility, trust in online stores, and sense of security. The study’s context, which compared biometrics with credit card authentication, offered theoretical evidence highlighting the significance of individuals’ perceptions, apprehensions, and beliefs regarding the adoption of biometrics in electronic payments. This study also had managerial implications, illuminating the importance of understanding individuals’ perceptions and concerns regarding secure authentication methods. Furthermore, the study underscored the necessity of implementing biometric authentication for electronic payments to address these concerns effectively.

In 2018, Chaudhry et al. [19] employed the random oracle model to assess the proposed scheme. This analysis confirmed that the scheme exhibited strong resilience against various potential attacks, with particular emphasis on guarding against smart card theft attacks and user anonymity violation attacks. The validity of the analysis was reinforced using the ProVerif automated software application. Furthermore, the analysis demonstrated that the proposed scheme boasted computational efficiency superior to that of Mir and Nikooghadam’s scheme.

In 2021, Tsai and Su [20] presented a novel approach for authenticating online banking customers and transactions were introduced, which involved the utilization of a hash-based multi-server authentication scheme alongside a smart card. The system’s design offered robust security attributes while minimizing maintenance expenses for financial institutions operating Internet banking services. This innovative mechanism was seamlessly integrated into established banking frameworks due to its compatibility with a tailored interface, making it an adaptable solution for enhanced security of Internet banking applications.

In 2019, Das et al. [21] developed an authentication method for ATMs, and employed a finger vein authentication technique, while a combined strategy encompassing steganography and cryptography was employed for transferring information, specifically finger vein images. The finger vein authentication system was realized through a fusion of image processing and machine learning techniques. As for the transfer of finger vein images, a hybrid solution involved lightweight cryptography and the variable MSB-LSB steganography algorithm was proposed. In the finger-vein authentication experiment, the results highlighted the efficacy of the proposed classification methods.

In 2015, Nagaraju and Parthiban [22] described the Multi-Factor Biometric Fingerprint Authentication (MFA) approach as systematically devised, offering a highly secure means of verifying the identities of remote users. A noteworthy feature of this approach was that user authentication details remained undisclosed to both the bank and cloud authentication servers, enabling these servers to perform remote user authentication without accessing the credentials. Building upon this researched framework, a privacy-preserving gateway was developed. This gateway employed tokenization and data anonymization techniques to obscure and desensitize customers’ account information. Notably, this approach maintained the original data format at various levels of the database management systems, rendering the data meaningless to unauthorized parties while remaining accessible to the rightful owner. Alongside the efficient MFA design, extensive experimental outcomes underscored the practicality and efficacy of the privacy protection gateway.

In 2022, Gavaskar et al. [23] prevented a security feature to counteract theft attempts involving the use of chloroform. A relay mechanism was incorporated, enabling the remote release of chloroform inside the ATM in response to a potential threat. Authorized personnel was initiated this action using a mobile app. To monitor and oversee the ATM’s internal activity, an ESP32-based camera was deployed, providing live video coverage. This camera not only recorded events but also transmitted video footage from inside the ATM. Additionally, the system tracked the ATM’s geographical location using GPS, providing latitude and longitude coordinates. The entire system was integrated with the Blynk mobile application. The microcontroller collected data from sensors and the GPS module. This information was then transmitted to the Blynk application. Through the Blynk application, authorized personnel was interact with the system. They had the capability to manipulate the relays, thus controlling the connected devices by toggling them ON or OFF. This versatile setup was to find applications in real-time scenarios like banking institutions, residences, and industrial settings.

In 2018, Sujatha and Chilambuchelvan [24] designed a sophisticated multimodal biometric algorithm, palm print, combining iris, signature recognition, and face techniques using an encoded discrete wavelet transform for comprehensive authentication and image analysis. This approach employed a multi-level wavelet-based fusion method, effectively integrating and encoding the biometric traits into a single composite image to facilitate matching decisions. The integration process was not only minimized memory requirements but also enhanced recognition accuracy and Equal Error Rate (ERR) compared to standalone biometric methods. The fusion and reconstruction algorithms were adeptly designed to handle the complexity of the process, making them well-suited for a wide range of real-time applications.

In 2021, Sangeetha et al. [25] implemented a fingerprint authentication method within the ATM system was undertaken with the primary goal of enhancing safety and security for individuals while streamlining transaction processes. The inherent uniqueness of fingerprints for each person ensured a robust level of security. This approach eliminated concerns related to lost ATM cards and eliminated the need to carry physical cards at all times. In comparative evaluations of various technologies for ATM security, fingerprint technology emerged as a superior and more secure option. These factors collectively established this mechanism as a convenient and highly secure transaction method, fostering a seamless interaction between users and ATM machines. Undoubtedly, this technology marked a cutting-edge advancement in electronic cash transactions.

Numerous studies regarding ATM transactions are given in Table 1. Fingerprint-based authentication [18] enhances the system performance of doing the transaction. But, the external validity becomes less which can mislead the authentication process. Key agreement [19] is more resistible to certain security threats by using the symmetric key-based authentication. Still, it incurs the computation overhead or complexity that affects the system integrity. Two-factor authentication [20] enhances the security level by considering the two different factor processes. However, it requires more training process that makes the system more intricate. SVM [21] offers higher classification accuracy and it can also reduce the time required for execution. On the other hand, it is less sensitive to noise and it also contains inadequate pixel information. MFA [22] performs better for data extraction, symmetric encryption, and decryption. As it considers the cloud datasets, the system cannot able to handle the massive collection of resources. IoT-based system [23] yields superior results in monitoring ATM transactions and their security. On the other hand, it suggests including biometric-based authentication for a better process. DWT [24] achieves the desired authentication value and it also offers multimodal biometric information. Nevertheless, the behavior-related traits, multi-sensor and multi-instance are critical to play in this process. Fingerprint-based verification [25] becomes effortless to avail for better verification by ATM system, but, it entails hardware complexity. These help to motivate to design of a well-developed model for ATM transactions.

Table 1
Features and challenges of traditional multimodal biometric-based authentication for ATM transaction

Methodology	Features	Challenges	Author [citation]
Fingerprint-based authentication	It enhances the system performance of doing the transaction.	The external validity becomes less which can mislead the authentication process.	Ogbanufe and Kim [18]
Key agreement	It is more resistible to certain security threats by using the symmetric key-based authentication.	It incurs the computation overhead or complexity that affects the system integrity.	Chaudhry et al. [19]
Two-factor authentication	It enhances the security level by considering the two different factor processes.	It requires more training process that makes the system more intricate.	Tsai and Su [20]
SVM	It offers higher classification accuracy. It can also reduce the time required for execution.	It is less sensitive to noise and it also contains inadequate pixel information.	Das et al. [19]
MFA	It performs better for data extraction, symmetric encryption, and decryption.	As it considers the cloud datasets, the system cannot able to handle the massive collection of resources.	Nagaraju and Parthiban [22]
IoT-based system	It yields superior results in monitoring the ATM transactions and its security.	It suggests including biometric-based authentication for a better process.	Gavaskar et al. [23]
DWT	It achieves the desired authentication value. It also offers multimodal biometric information	Behavior-related traits, multi-sensor, and multi-instance are critical to play in this process.	Sujatha and Chilambuchelvan [24]
Fingerprint-based verification	It becomes effortless to avail for better verification by ATM system.	It entails the hardware complexity.	Sangeetha et al.

Figure 1.

Architectural view of recommended security framework on ATM transactions based on Multi-modal biometric authentication using deep networks.

3. Proposed model description and multiple biometric-based data collection for authentication for ATM transaction

3.1 Proposed secured authentication model for ATM transaction

ATMs manage physical currency and sensitive financial data, underscoring the critical importance of robust security protocols to forestall unauthorized access and fraudulent transactions. The potential repercussions, encompassing financial harm to individuals and financial institutions, underscore the urgency of establishing formidable security measures for ATM transactions. A plethora of security frameworks have been developed to elevate user protection. While conventional ATM transaction methods have undoubtedly offered convenience, they also bring forth several shortcomings relating to security, accessibility, and potential vulnerabilities. Illicit actors may resort to tactics like shoulder surfing to surreptitiously acquire PINs, granting them unauthorized access to accounts and the capability to execute deceitful transactions. Furthermore, criminals might employ card cloning by pilfering card information during ATM transactions [26], enabling them to carry out unauthorized purchases. The nefarious practice of ATM card skimming involves affixing devices to ATMs [27] to covertly gather card data. To counteract these security risks, this research endeavors to harness biometric attributes such as fingerprints, iris scans, speech recognition, and facial recognition. This biometrics is chosen for their distinctiveness, rendering them exceedingly arduous to replicate. By incorporating these biometric elements, the research seeks to establish confidentiality, integrity, and availability of financial services and data, thereby ensuring a secure and dependable platform for conducting day-to-day financial operations. Figure 1 provides an illustration of the Proposed Security Framework for ATM transactions, leveraging multi-modal biometric authentication through deep networks.

The proposed system consists of many phases, the primary goal of the system is to enhance the security of ATM transactions by implementing a novel multimodal biometric-based authentication approach. This approach combines multiple biometric factors (face, retina, fingerprint, and spectrogram) to ensure a higher level of security and accuracy, reducing the risks associated with traditional card and PIN methods. To propose MEB7-AM is designed to efficiently process and analyze the biometric information from various sources. This mechanism ensures that each channel’s data contributes effectively to the authentication process, thereby increasing accuracy and reducing false positives. After successful multimodal authentication, the system streamlines the transaction process. Users only need to enter a single PIN to select the desired bank. Once the bank is selected and proper credentials are provided, the ATM machine facilitates the withdrawal of money seamlessly. The proposed system aims to assess its efficacy using a comprehensive set of performance measures. By comparing the results with existing authentication methodologies, the system’s accuracy, speed, and reliability can be quantified, demonstrating its superiority and effectiveness.

4. Authentication for ATM transactions Using Multichannel EfficientNet B7 with attention mechanism

4.1 Multichannel EfficientNet B7

Using a Multichannel EfficientNet B7, the ATM authentication system sounds like an advanced and secure approach. EfficientNet B7 is a convolutional neural network architecture that is known for its efficiency and effectiveness in image classification tasks. Combining it with a multichannel approach can enhance its capabilities for ATM authentication. The explanation of Multichannel EfficientNet B7 is explained below

Multichannel [28]: In the context of deep learning, “multichannel” typically refers to the use of multiple input channels in a neural network model. These input channels represent different sources or types of data that are fed into the network simultaneously. Each input channel contains distinct information, and the network processes these channels collectively to make predictions or perform tasks. The concept of multichannel in deep learning enhances the model’s ability to learn and extract features from diverse data sources, ultimately improving its performance on various tasks. It’s commonly used in tasks like image classification, object detection, and speech recognition, where different aspects of the input data can provide complementary information for better predictions or analysis.

EfficientNet B7:

EfficientNet B7 [29], which belongs to the EfficientNet model family, stands as a Deep Convolutional Neural Network (DCNN) architecture meticulously crafted to find an optimal balance between precision and computational efficiency. This inherent adaptability renders it particularly well-suited for an array of computer vision tasks, encompassing image classification and object detection. The term “Efficient” in EfficientNet encapsulates the approach of methodically scaling the model’s dimensions to attain heightened performance gains without imposing significant increments in computational requirements. This scaling is achieved through a comprehensive approach that concurrently adjusts the network’s depth, width, and resolution while also governing parameters and computational intricacies.

In particular, EfficientNet B7 stands as one of the more substantial and potent variants within the EfficientNet lineup. Compared to smaller iterations like EfficientNet B0 or B1, EfficientNet B7 possesses more parameters and demands higher computational resources. This heightened capacity equips EfficientNet B7 to tackle more intricate tasks and datasets, delivering elevated levels of accuracy. The architecture of EfficientNet B7 retains the foundational building blocks characteristic of other EfficientNet models, encompassing squeeze-and-excitation blocks, judicious dimension scaling, and depthwise separable convolutions. EfficientNet includes versions from B0 to B7 where each with diverse constraints. The segmentation issues were successfully resolved by the transfer learning model [30].

Given its larger dimensions and capabilities, EfficientNet B7 finds suitability in scenarios necessitating remarkable accuracy. It shines in tasks where sufficient computational resources are accessible to accommodate its augmented model complexity. Nevertheless, it’s essential to acknowledge that leveraging such expansive models could entail greater memory usage, processing potency, and extended training durations. Figure 2 shows the architecture diagram for Multichannel EfficientNet B7.

Figure 2.

Architecture diagram for Multichannel EfficientNet B7.

4.2 Suggested Multichannel EfficientNet B7 with attention mechanism for authentication

One critical application is in the realm of ATM transactions, where user identity verification plays a pivotal role in safeguarding financial transactions and personal information. To address the challenges of enhancing both security and efficiency, a cutting-edge solution called the MEB7-AM for ATM Transaction Authentication has been developed. The attention mechanism is explained below.

Attention Mechanism [31]: Within this framework, the attention vector, denoted as $m$ , emerges through the amalgamation of three crucial components: The query vector $l$ , The value matrix $s$ and The key matrix $u$ .

Expanding upon this, the keys serve as a pivotal function for the value matrix, as delineated in Eq. (1).

$\displaystyle E=\tanh({sp^{l}})$ (1)

In this context, the parameterization is denoted as $p^{l}$ , which carries critical importance across individual time steps. This importance is evaluated based on the dot product’s magnitude between each key vector and the corresponding query vector, termed as $u\in\Re^{B_{a}}$ , with reference to the attention dimension hyperparameter $B_{a}$ .

These magnitudes are subsequently employed in both weight determination and the computation of the weighted sum of value vectors. This computation $k$ is succinctly expressed in Eqs (2) and (3).

$\displaystyle y=S_{t}^{\max}({KL^{t}})$ (2) $\displaystyle l=ds$ (3)

Furthermore, the attention mechanism has introduced a mechanism for enabling the direct selection of detection, contingent upon subsets of pertinent tokens. This mechanism is facilitated through the function involving the concatenation of $r_{(u)}$ and the attention vector denoted as $r_{({s-1})}$ . This attention vector represents the weighted sum over hidden states and is delineated in Eq. (4).

$\displaystyle r_{(u)}=S_{t}^{\max}({[{r_{({s-1})}gy_{({s-1})}}]W+qi})$ (4)

MEB7-AM authentication system represents a groundbreaking approach to enhancing the security and efficiency of ATM transactions. Here the collected data $PF_{a}^{\textit{input}}$ , $AF_{b}^{\textit{input}}$ , $ER_{c}^{\textit{input}}$ , $SI_{d}^{\textit{input}}$ is inserted into the multichannel and the output of the multichannel is fed as input to the attention mechanism, finally the outcome of the attention mechanism is given as input to the of EfficientNet B7, by leveraging the power of deep learning, multichannel fusion, and attention mechanisms, this system provides a robust and intelligent solution that can accurately verify user identities in real-time, thereby ensuring a secure and seamless ATM experience for users while mitigating the risks associated with unauthorized access and fraudulent activities. Figure 3 shows the architecture diagram of the proposed MEB7-AM for ATM authentication.

Figure 3.

Architecture diagram of proposed MEB7-AM for ATM authentication.

4.3 Procedure for authentication

The primary goal of the system is to enhance the security of ATM transactions by implementing a novel multimodal biometric-based authentication approach. This approach combines multiple biometric factors face, retina, fingerprint, and spectrogram to ensure a higher level of security and accuracy, reducing the risks associated with traditional card and PIN methods. This helps to prevent various fraudulent activities, such as card theft and PIN compromise.

The system aims to integrate multiple banks’ access into a single ATM card provided by a Multi-Bank Provider. This eliminates the need for individuals to carry multiple cards for different banks, simplifying the user experience and making ATM transactions more convenient.

The proposed MEB7-AM is designed to efficiently process and analyze biometric information from various sources (face, retina, fingerprint, and spectrogram). This mechanism ensures that each channel’s data contributes effectively to the authentication process, thereby increasing accuracy and reducing false positives. After successful multimodal authentication, the system streamlines the transaction process.

Users only need to enter a single PIN to select the desired bank. Once the bank is selected and proper credentials are provided, the ATM machine facilitates the withdrawal of money seamlessly. The proposed system aims to assess its efficacy using a comprehensive set of performance measures. Figure 4 shows the flowchart for the proposed authentication for the ATM transaction system.

Figure 4.

Flowchart for proposed authentication for ATM transaction system.

5. Results and discussion

5.1 Simulation setup

Python was used for the implementation of Authentication for the ATM Transaction system, and the analysis will be carried out. Some of the classifiers like LSTM [32], RNN [33], ResNet [34], CNN [35], and EfficientNet B7 [29] were assimilated in this newly proposed model. The loss function used for the deep learning models was discussed as follows. The loss function used for the LSTM and RNN model is a mean squared error. Accordingly, the loss functions for the ResNet and EfficientNet B7 models are binary cross-entropy and categorical cross-entropy. The simulation parameters of the models like LSTM, RNN, ResNet, CNN, and EfficientNet B7 are shown in Table 2.

Table 2
Simulation parameters of the models like LSTM, RNN, ResNet, CNN, and EfficientNet B7

Methods	Parameters	Values
LSTM	Hidden neuron count	[5, 255]
	Epochs	[50, 100]
	Batch-size	[4, 256]
	Steps-per-epoch	[300, 1000]
RNN	Epochs	[50, 100]
	Batch-size	[4, 256]
	Optimizer	[0, 4]
ResNet	Activation	[0, 4]
	Hidden neuron count	[5, 255]
CNN	Epochs	[50, 100]
	Activation	[0, 4]
	Optimizer	[0, 4]
EfficientNet B7	Epochs	[50, 100]
	Steps-per-epoch	[300, 1000]
	Activation	[0, 4]
	Optimizer	[0, 4]
	Hidden neuron count	[5,255]

5.2 Multiple biometric information

Within the scope of this research, biometric data from bank users is gathered and employed. Specifically, samples of biometric information encompassing fingerprints, facial features, retinal patterns, and speech characteristics are compiled from an online dataset. The subsequent section provides a concise overview of this data collection process.

Fingerprint Data: The biometric data pertaining to the fingerprints of bank users is procured from a dataset known as the “Sokoto Coventry Fingerprint Dataset (SOCOFing).” This dataset can be accessed through the URL “https://www.kaggle.com/code/jamalahmedkhan/fingerprint-matching/input,” as of the date August 16, 2023. The input data is categorized within the “Altered and Real” subfolders situated under the main SOCO Fing directory. Designed specifically for research objectives, this dataset encompasses a collection of 6,000 fingerprint images extracted from 600 distinct African subjects.

Face Data: The facial data is obtained from the “Georgia Tech face database,” accessible via the link “http://www.anefian.com/research/face_reco.htm” as of August 16, 2023. This database occupies a storage space of 128 MB and comprises photographic images of 50 individuals. The images are presented in JPEG format and feature a varied background setting. Each facial image possesses dimensions of 150 $\times$ 150 pixels.

Retina Data: The dataset containing retina image data is sourced from the “Retinal Vessel Segmentation” dataset, which was accessed on August 16, 2023; via the URL “https://www.kaggle.com/datasets/ipythonx/ retinal-vessel-segmentation” This dataset encompasses a collection of images capturing retinal vessels.

Speech Signal: The speech signal examples are collected from the “TIMIT Corpus Sample (LDC93S1)” dataset, acquired from the link “https://www.kaggle.com/datasets/nltkdata/timitcorpus” on August 16, 2023. Designed to provide speech data for the advancement of automatic speech recognition systems and phonetic research, this dataset consists of a total of 130 sentences and 160 voice recordings.

The fingerprint input is symbolized by the term $\textit{PF}_{a}^{\textit{input}}$ , and the face image input is denoted by $\textit{AF}_{b}^{\textit{input}}$ . The retina input is represented by $\textit{ER}_{c}^{\textit{input}}$ , and the speech signal input is indicated by $\textit{SD}_{d}^{\textit{input}}$ .

In the procedure of recognizing speech signals, noise reduction is implemented to eliminate extraneous sounds from the speech signal. Furthermore, the transformation of the signal into spectrogram images $\textit{SI}_{d}^{\textit{input}}$ is achieved through Short-Time Fourier Transform (STFT) [36]. This technique presents an enhanced and precise method for confirming identity, thereby reducing the likelihood of incorrect identification or unauthorized entry into accounts. Additionally, it furnishes authentication in real-time, bolstering the security of transactions and guaranteeing that only sanctioned users can commence and finalize transactions. The data size is (100, 4), and also 4 features were attained for the implementation process. The whole data is divided into two sections like training and testing where 75% of the data is given to the training stage and also 25% of the data is fed to the testing stage. Figure 5 shows the sample Images of the Biometric details of the bank users.

Figure 5.

Samples of face, finger, retina, speech.

5.3 Performance measures

Performance measures are below.

$\displaystyle\textit{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$ (5) $\displaystyle\textit{precision}=\frac{TP}{TP+FP}$ (6) $\displaystyle\text{Negative Predictive Value}\quad\textit{NPV}=\frac{TN}{TN+FN}$ (7) $\displaystyle\text{False Positive Rate}\quad\textit{FPR}=\frac{FP}{FP+TN}$ (8) $\displaystyle\textit{sensitivity}=\frac{TP}{TP+FN}$ (9) $\displaystyle\textit{specificity}=\frac{TN}{TN+FP}$ (10) $\displaystyle\text{False Negative Rate}\quad\textit{fnr}=\frac{FN}{FN+TP}$ (11) $\displaystyle\text{False Discovery Rate}\quad\textit{fdr}=\frac{FP}{FP+TP}$ (12) $\displaystyle\textit{f1-score}=\frac{2TP}{2TP+FP+FN}$ (13)

where the terms $T P$ , $F N$ , $F P$ and $T N$ provide the respective values for true positive, false negative, false positive, and true negative.

Figure 6.

Confusion matrix analysis of proposed authentication for ATM transaction model compared over the classical model.

5.4 Evaluation of the confusion matrix for the created model for authentication for an ATM transaction system

The system’s evaluation involved the application of a confusion matrix for the MEB7-AM model, and the results are illustrated in Fig. 6. The evaluation criteria, including metrics such as accuracy, were assessed using this confusion matrix. Figure 6 shows the confusion matrix for ATM cards with PIN, the values range from 0 to 400, potentially denoting the spectrum of data points or categories under classification. This range signifies how accurately the system’s predictions align with the true labels. The proposed model exhibits proficient categorization, indicating its robust performance in terms of categorization capabilities, as indicated by the metrics depicted in the visual representation.

5.5 Evaluation of the ROC analysis for the created model for authentication for the ATM transaction system

The Receiver operating characteristic (ROC) curve serves as a visual tool for evaluating and gauging the accuracy of classification models, specifically in scenarios involving binary classification tasks. It visually presents the interplay between the true positive rate and the false positive rate, which pivots around the shifting threshold of a classification model. The ROC curve is employed to compare the efficacy of a proposed model against an established classifier, as depicted in Fig. 7. This comparison entails a comprehensive exploration of potential models, encompassing the spectrum of false and true positive measures spanning from 0 to 1. In essence, this evaluation approach involves analyzing how the models’ performance characteristics unfold across different threshold settings, illuminating the dynamics of their true positive and false positive rates.

Figure 7.

Receiver operating characteristic (ROC) analysis of proposed authentication for ATM transaction model compared over the classical model.

Table 3

Overall determination for the proposed ATM authentication model regarding the classifiers

Measures	LSTM [32]		RNN [33]		ResNet [34]		EfficientNet B7 [35]		MEB7-AM
Fingerprint
Accuracy	86	.7	88	.8	91	.3	93	.4	93	.9
Recall	87	.71267	88	.65784	91	.30435	93	.38374	93	.95085
Specificity	85	.56263	88	.95966	91	.29512	93	.41826	93	.84289
Precision	87	.21805	90	.01919	92	.17557	94	.09524	94	.48669
False positive rate (FPR)	14	.43737	11	.04034	8	.704883	6	.581741	6	.157113
False negative rate (FNR)	12	.28733	11	.34216	8	.695652	6	.616257	6	.049149
Negative predictive value (NPV)	86	.11111	87	.4739	90	.33613	92	.63158	93	.24895
False discovery rate (FDR)	12	.78195	9	.980806	7	.824427	5	.904762	5	.513308
F1-score	87	.46466	89	.33333	91	.73789	93	.73814	94	.21801
MCC (matthews correlation coefficient)	0	.733022	0	.775553	0	.825556	0	.867644	0	.877647
Face recognition
Accuracy	88	.4	90	.6	91	.9	94		94	.7
Recall	88	.61004	90	.92664	91	.89189	94	.01544	94	.78764
Specificity	88	.17427	90	.24896	91	.90871	93	.9834	94	.60581
Precision	88	.95349	90	.92664	92	.42718	94	.37984	94	.97099
False positive rate (FPR)	11	.82573	9	.751037	8	.091286	6	.016598	5	.394191
False negative rate (FNR)	11	.38996	9	.073359	8	.108108	5	.984556	5	.212355
Negative predictive value (NPV)	87	.80992	90	.24896	91	.34021	93	.59504	94	.40994
False discovery rate (FDR)	11	.04651	9	.073359	7	.572816	5	.620155	5	.029014
F1-score	88	.78143	90	.92664	92	.15876	94	.19729	94	.87923
MCC (matthews correlation coefficient)	0	.767739	0	.811756	0	.83784	0	.879869	0	.893872
Retina
Accuracy	87	.7	89	.6	88		91	.5	95	.9
Recall	87	.2	89	.4	88		91	.6	96
Specificity	88	.2	89	.8	88		91	.4	95	.8
Precision	88	.08081	89	.75904	88		91	.41717	95	.80838
False positive rate (FPR)	11	.8	10	.2	12		8	.6	4	.2
False negative rate (FNR)	12	.8	10	.6	12		8	.4	4
Negative predictive value (NPV)	87	.32673	89	.44223	88		91	.58317	95	.99198
False discovery rate (FDR)	11	.91919	10	.24096	12		8	.582834	4	.191617
F1-score	87	.63819	89	.57916	88		91	.50849	95	.9041
MCC (matthews correlation coefficient)	0	.754038	0	.792006	0	.76	0	.830002	0	.918002
Speech
Accuracy	89	.5	91	.4	93	.3	94	.7	95	.4
Recall	89	.31452	91	.53226	93	.14516	94	.55645	95	.3629
Specificity	89	.68254	91	.26984	93	.45238	94	.84127	95	.43651
Precision	89	.49495	91	.16466	93	.33333	94	.74747	95	.3629
False positive rate (FPR)	10	.31746	8	.730159	6	.547619	5	.15873	4	.563492
False negative rate (FNR)	10	.68548	8	.467742	6	.854839	5	.443548	4	.637097
Negative predictive value (NPV)	89	.50495	91	.63347	93	.26733	94	.65347	95	.43651
False discovery rate (FDR)	10	.50505	8	.835341	6	.666667	5	.252525	4	.637097
F1-score	89	.40464	91	.34809	93	.23915	94	.65187	95	.3629
MCC (matthews correlation coefficient)	0	.754038	0	.792006	0	.76	0	.830002	0	.918002

Figure 8.

K-fold evaluation of the designed model.

Figure 8.

Continued.

5.6 Performance analysis of the proposed authentication for ATM transaction method

The effectiveness of the suggested approach is evaluated using various classifiers are shown in Fig. 8. Here, the entire dataset is divided into 5 sets. While considering the 100% data for processing at the 1-fold includes 1–20, 2-fold contains 21–40, 3-fold includes 41–60, 4-fold contains 61–80, and 4-fold which includes 81–100. Initially, if the 1-fold analysis has been considered, where testing is made on the 1^st set and then the other sets of data are utilized for the training stage. Based on this procedure the whole implementation is done. With the accuracy of Fingerprint, the proposed system outperforms other methods with notable differences: The accuracy of the suggested model stands at 2.3%, 4.4%, 7%, 6.7%, and 6.5% in relation to LSTM, RNN, ResNet, CNN, and EfficientNet B7 respectively. This accuracy is comparatively lower than that of the proposed MEB7-AM model. This emphasizes that a higher accuracy value contributes to the proficient identification of an ATM authentication system, making it a more effective tool for authentication.

5.7 Overall validation of the proposed ATM authentication model using diverse algorithms and classifiers

Table 3 illustrates the evaluation of the proposed ATM authentication model across various classifier models. When compared to alternative classifiers such as LSTM, RNN, ResNet, CNN, and EfficientNet B7, the MEB7-AM model demonstrates an MCC (Matthews Correlation Coefficient) value that respectively 18%, 14%, 11%, 3.4%, and 6% higher. This outcome substantiates the efficacy of the provided ATM authentication model.

6. Conclusion

Biometric-based Authentication for ATM transactions by the MEB7-AM model was carried out using Python. The proposed system comprised several phases, with its core objective being the enhancement of ATM transaction security through the implementation of an innovative multimodal biometric-based authentication approach. This method synergized various biometric factors such as face, retina, fingerprint, and spectrogram to ensure a heightened level of security and precision, mitigating the vulnerabilities inherent in conventional card and PIN techniques. The introduction of MEB7-AM aims to optimize the processing and evaluation of biometric data from diverse sources. This mechanism guaranteed the effective integration of data from each channel, thus augmenting accuracy and diminishing instances of false positives. Following successful multimodal authentication, the system optimized the transaction procedure, requiring users to input a sole PIN for bank selection. Upon selecting the bank and providing valid credentials, the ATM machine facilitated seamless money withdrawal. The proposed system’s effectiveness was evaluated through a comprehensive set of performance metrics. By juxtaposing outcomes against established authentication methods, the system’s superiority and efficacy in terms of accuracy, speed, and reliability can be ascertained. The accuracy of the suggested model stands at 7.3%, 8.4%, 7%, 3.7%, and 7.5% in relation to LSTM, RNN, ResNet, CNN, and EfficientNet B7 respectively. This accuracy was comparatively lower than that of the proposed MEB7-AM model. This emphasized that a higher accuracy value contributes to the proficient identification of an ATM authentication system, making it a more effective tool for authentication. Biometric data is highly sensitive, and ensuring its privacy and security is a critical challenge. Protecting biometric information from theft, unauthorized access, and potential breaches is essential. The proposed biometric-based authentication system using the MEB7-AM model presents exciting opportunities to enhance ATM transaction security. However, it also faces challenges related to data privacy, robustness, and practical implementation. Addressing these challenges and exploring future advancements can pave the way for more secure and user-friendly authentication solutions in the financial sector.

Footnotes

Author’s Bios

	M. Ravi Prasad is working as Sr. Assistant Professor in SV college of Engineering from July-2019 in the Department of CSE. Also worked as HOD of MCA and Incharge of Examinations at Sri Sai College of IT and Management from July 2011 to 2019.
	N. Thillaiarasu is currently working as an Associate Professor from January 2021 in the School of Computing and Information Technology, REVA University, Bengaluru. He has also served as an Assistant Professor at Galgotias University, Greater Noida, from July 2019 to December 2020. He worked from April 2012 to June 2019 as an Assistant Professor in the Department of Computer Science and Engineering, SNS College of Engineering, Coimbatore. He obtained his B.E., in Computer Science and Engineering from Selvam College of Technology in 2010 and his M.E., in Software Engineering from Anna University Regional Center, Coimbatore, in 2012. He received his Ph.D., Degree from Anna University, Chennai, in 2019, and he has published more than 40 research papers in refereed, Springer, and IEEE Xplore conferences. He has organized several workshops, summer internships, and expert lectures for students. He has worked as a session chair, conference steering committee member, editorial board member, and reviewer in the Springer Journal and IEEE Conferences. He has edited books for Bentham Science and
CRC Press, Taylor & Francis. His area of interest include Cloud Computing, Security, IoT, and Machine Learning. ORCID: 0009-0006-8267-8393.

References

Shimal

and Debbarma

, Designing a Biometric Strategy (Fingerprint) Measure for Enhancing ATM Security in Indian E-Banking System, Computer Science, 2011.

Assaad

F.S.

and Serpen

, Transformation-based score fusion algorithm for multi-modal biometric user authentication through ensemble classification, Procedia Computer Science 61 (2015), 410–415.

Breebaart

Buhan

Groot

and Kelkboom

, Evaluation of a template protection approach to integrate fingerprint biometrics in a PIN-based payment infrastructure, Electronic Commerce Research and Applications 10(6) (2011), 605–614.

William

Pawar

A.B.

Jawale

M.A.

Badholia

and Verma

, Energy efficient framework to implement next generation network protocol using ATM technology, Measurement: Sensors 24 (2022).

Purohit

and Ajmera

P.K.

, Optimal feature level fusion for secured human authentication in multimodal biometric system, Machine Vision and Applications 32(24) (2021).

Yoon

and Yoo

K.Y.

, Robust biometrics-based multi-server authentication with key agreement scheme for smart cards on elliptic curve cryptosystem, The Journal of Supercomputing 63 (2013), 235–255.

Lee

and Kwak

, Construction of a secure two-factor user authentication system using fingerprint information and password, Journal of Intelligent Manufacturing 25 (2014), 217–230.

Alhothaily

Alrawais

Cheng

and Bie

, A novel verification method for payment card systems, Personal and Ubiquitous Computing 19 (2015), 1145–1156.

Chang

Y.F.

Tai

W.L.

and Chang

H.C.

, Untraceable dynamic-identity-based remote user authentication scheme with verifiable password update, International Journal of Communication System 27(11) (2014), 3430–3440.

10.

Wang

and Chu

, Anonymous two-factor authentication in distributed systems: certain goals are beyond attainment, IEEE Trans Dependable Secure Computer, 2014.

11.

Sudharsan

Kumar

V.D.A.

Venkatesan

Sathyapreiya

and Saranya

, Two three step authentication in ATM machine to transfer money and for voting application, Procedia Computer Science 165 (2019), 300–306.

12.

Ogbanufe

and Kim

D.J.

, Comparing fingerprint-based biometrics authentication versus traditional authentication methods for e-payment, Decision Support Systems 106 (2018), 1–14.

13.

Khan

M.K.

and Kumari

, An improved biometrics-based remote user authentication scheme with user anonymity, Biomed Res Int, 2013.

14.

Noh

K.S.

, A study on the authentication and security of financial settlement using the finger vein technology in wireless internet environment, Wireless Personal Communications 89 (2016), 761–775.

15.

Graevenitz

G.A.

, Biometric authentication in relation to payment systems and ATMs, Datenschutz und Datensicherheit – DuD 31 (2007), 681–683.

16.

Chatterjee

and De

, A novel multi-server authentication scheme for e-commerce applications using smart card, Wireless Personal Communications 91 (2016), 293–312.

17.

Shingala

Patel

and Doshi

, An improve three factor remote user authentication scheme using smart card, Wireless Personal Communications 99 (2018), 227–251.

18.

Ogbanufe

and Kim

D.J.

, Comparing fingerprint-based biometrics authentication versus traditional authentication methods for e-payment, Decision Support Systems 106 (2018), 1–14.

19.

Chaudhry

S.A.

Naqvi

and Khan

M.K.

, An enhanced lightweight anonymous biometric-based authentication scheme for TMIS, Multimedia Tools and Applications 77 (2018), 5503–5524.

20.

Tsai

C.H.

and Su

P.C.

, The application of multi-server authentication scheme in internet banking transaction environments, Information Systems and e-Business Management 19 (2021), 77–105.

21.

Das

Singh

Gupta

Banerjee

Mohiuddin

and Tiwary

, Design and implementation of secure ATM system using machine learning and crypto-stego methodology, SN Applied Sciences 1 (2019), 976.

22.

Nagaraju

and Parthiban

, Trusted framework for online banking in public cloud using multi-factor authentication and privacy protection gateway, Journal of Cloud Computing 4 (2015), 22.

23.

Gavaskar

Ragupathy

U.S.

Elango

Ramyadevi

and Preethi

, A novel design and implementation of IoT based real-time ATM surveillance and security system, Advances in Computational Intelligence 2(1) (2022).

24.

Sujatha

and Chilambuchelvan

, Multimodal biometric authentication algorithm using iris, palm print, face and signature with encoded DWT, Wireless Personal Communications 99 (2018), 23–34.

25.

Sangeetha

Kumaraguru

Akshay

and Kanishka

, Biometric based Fingerprint Verification System for ATM machines, in: Journal of Physics: Conference Series, 2021.

26.

Mandalapu

Deepa

V.D.

Raj

L.D.

and Dev

J.A.

, An NFC featured three level authentication system for tenable transaction and abridgment of ATM card blocking intricacies, in: 2015 International Conference and Workshop on Computing and Communication (IEMCON), 2015, pp. 1–6.

27.

Nesakumar

A.D.

Arthi

Lahari

Geetha

Pavithra

K.N.

and Mugilan

, Smart ATM Card for Multiple Bank Accounts, in: 2022 International Interdisciplinary Humanitarian Conference for Sustainability (IIHC), 2022, pp. 1228–1232.

28.

Suwa

Yamamoto

Tsuchida

Nakamura

Wakayama

and Hara

, Image-based target detection and radial velocity estimation methods for multichannel SAR-GMTI, IEEE Transactions on Geoscience and Remote Sensing 55(3) (2017), 1325–1338.

29.

Selim

Elkabani

and Abdou

M.A.

, Students Engagement Level Detection in Online e-Learning Using Hybrid EfficientNetB7 Together With TCN, LSTM, and Bi-LSTM, in IEEE Access 10 (2022), 99573–99583.

30.

Khalil

M.I.

Tehsin

Humayun

Jhanjhi

N.Z.

and AlZain

M.A.

, Multi-scale network for thoracic organs segmentation, Computers, Materials & Continua 70(2) (2022), 3252–3265.

31.

Feng

Zhang

and Song

, Unrestricted attention may not be all you need-masked attention mechanism focuses better on relevant parts in aspect-based sentiment analysis, IEEE Access 10 (2022), 8518–8528.

32.

Dai

and Li

, Modeling vehicle interactions via modified LSTM models for trajectory prediction, IEEE Access 7 (2019), 38287–38296.

33.

Zhang

and Liu

, MM-RNN: A Multimodal RNN for Precipitation Nowcasting, IEEE Transactions on Geoscience and Remote Sensing 61(4101914) (2023), 1–14.

34.

Zhu

, Juggler-ResNet: A flexible and high-speed ResNet optimization method for intrusion detection system in software-defined industrial networks, IEEE Transactions on Industrial Informatics 18(6) (2022), 4224–4233.

35.

Kollias

and Zafeiriou

, Exploiting Multi-CNN Features in CNN-RNN Based Dimensional Emotion Recognition on the OMG in-the-Wild Dataset, IEEE Transactions on Affective Computing 12(3) (2021), 595–606.

36.

Alon

Shor

Moshkin

Dagan

and Singer

, Envelope tracking power amplifier using short-time fourier transform, IEEE Microwave and Wireless Components Letters 31(6) (2021), 575–578.

Multichannel EfficientNet B7 with attention mechanism using multimodal biometric- based authentication for ATM transaction

Abstract

Keywords

1. Introduction

2. Literature review

Table 1 Features and challenges of traditional multimodal biometric-based authentication for ATM transaction

3.1 Proposed secured authentication model for ATM transaction

4. Authentication for ATM transactions Using Multichannel EfficientNet B7 with attention mechanism

4.1 Multichannel EfficientNet B7

EfficientNet B7:

5.1 Simulation setup

Table 2 Simulation parameters of the models like LSTM, RNN, ResNet, CNN, and EfficientNet B7

5.5 Evaluation of the ROC analysis for the created model for authentication for the ATM transaction system

5.7 Overall validation of the proposed ATM authentication model using diverse algorithms and classifiers

6. Conclusion

Footnotes

Author’s Bios

References

Table 1
Features and challenges of traditional multimodal biometric-based authentication for ATM transaction

Table 2
Simulation parameters of the models like LSTM, RNN, ResNet, CNN, and EfficientNet B7