
==== Front
Heliyon
Heliyon
Heliyon
2405-8440
Elsevier

S2405-8440(24)12421-8
10.1016/j.heliyon.2024.e36390
e36390
Research Article
A proposed biometric authentication hybrid approach using iris recognition for improving cloud security
El-Sofany Hosam helsofany@kku.edu.sa
a⁎
Bouallegue Belgacem ab
Abd El-Latif Yasser M. c
a College of Computer Science, King Khalid University, Abha, Kingdom of Saudi Arabia
b Electronics and Micro-Electronics Laboratory (E. μ. E. L), Faculty of Sciences of Monastir, University of Monastir, Monastir, Tunisia
c Faculty of Science, Ain Shams University, Cairo, Egypt
⁎ Corresponding author. helsofany@kku.edu.sa
16 8 2024
30 8 2024
16 8 2024
10 16 e363909 5 2024
14 8 2024
14 8 2024
© 2024 The Authors
2024
https://creativecommons.org/licenses/by-nc/4.0/ This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Biometric systems have gained attention as a more secure alternative to traditional authentication methods. However, these systems are not without their technical limitations. This paper presents a hybrid approach that combines edge detection and segmentation techniques to enhance the security of cloud systems. The proposed method uses iris recognition as a biometric paradigm, taking advantage of the iris' unique patterns. We performed feature extraction and classification using hamming distance (HD) and convolutional neural networks (CNN). We validated the experimental findings using various datasets, such as MMU, IITD, and CASIA Iris Interval V4. We compared the proposed method's results to previous research, demonstrating recognition rates of 99.50 % on MMU using CNN, 97.18 % on IITD using CNN, and 95.07 % on CASIA using HD. These results indicate that the proposed method outperforms other classifiers used in previous research, showcasing its effectiveness in improving cloud security services.

Keywords

Cloud security
Biometric authentication
Iris recognition
Edge detection
Segmentation
Hamming distance
Convolutional neural network
==== Body
pmc1 Introduction

Traditional identification processes like passwords or cards can be compromised if passwords are forgotten or cards are stolen (reviewer objection: The problem and solutions are not clearly stated here). On the other hand, biometric authentication systems offer a more reliable means of recognizing individuals without relying on what they possess or need to remember [1]. Biometric authentication involves analyzing distinctive biological features such as the iris, hands, voice, retina, face, fingerprints, DNA, or signature to validate a user's identification. While these features are more reliable than conventional security systems, they do have limitations. For instance, it is possible to copy signatures, contaminate fingerprints, and alter speech and face biometrics with aging, illness, scarring, or uneven lighting. Furthermore, prolonged exposure to infrared light can destroy the retina biometric [2].

In biometric authentication systems, the saved biometric data allows for the recognition and validation of a user's identity when accessing their account. Due to the uniqueness of this information, biometric authentication is considered considerably safer than conventional methods. As a result, biometric authentication has been widely implemented across various digital resources, such as computing devices, buildings, and offices. Among the different biometric authentication techniques, iris biometrics have garnered significant interest in recent years (reviewer objection: The absence of the problem and solutions is still not addressed here).

The iris is a circular entity located in the middle of the eye, between the sclera and the pupil. Despite being an internal organ, the cornea protects it from the outside environment and allows for external visualization. Previous studies have shown that the iris remains relatively unchanged throughout a person's lifetime, with only subtle alterations occurring during childhood [3].

Using iris recognition as the biometric model, the proposed work uses edge detection and segmentation to secure cloud systems. This approach uses HD and CNNs for feature extraction and classification. MMU, IITD, and CASIA Iris Interval V4 datasets corroborated empirical findings. Comparing the proposed approach to previous research, it achieved recognition rates of 99.50 % on MMU using CNN, 97.18 % on IITD using CNN, and 95.07 % on CASIA using HD. While biometric systems have gained attention as a more secure alternative to traditional authentication methods, these systems are not without their technical limitations. The proposed method outperforms other classifiers used in previous research, showcasing its effectiveness in improving cloud security services [[4], [5], [6], [7], [8]].

The proposed authentication system relies on iris personal recognition, which consists of four key steps, including (1) edge detection, (2) iris segmentation, (3) feature extraction, and (4) binary matching classification.

The system reads the input dataset and carries out edge detection, localization, segmentation, feature extraction, and classification. By combining segmentation and edge detection processes, the system can identify edges, borders, and semicircles in the image. We then use the resulting image for feature extraction and classification, which ultimately leads to improved accuracy.

The proposed method for iris recognition might not necessitate a normalization procedure. The proposed hybrid approach combines edge detection and segmentation techniques with iris recognition, utilizing feature extraction and classification methods such as Hamming distance and convolutional neural networks (CNN). Experiments showed that the suggested method had high recognition rates on different datasets, such as MMU, IITD, and CASIA Iris Interval V4, even without using any normalization steps. The recognition rates of 99.50 % on MMU using CNN, 97.18 % on IITD using CNN, and 95.07 % on CASIA using HD indicate the effectiveness of the proposed method in improving cloud security services. Therefore, based on the results, it can be inferred that the proposed hybrid approach for iris recognition may not require a normalization procedure to achieve high recognition rates and enhance cloud security.

Feature extraction and classification using Hamming distance (HD) and convolutional neural networks (CNNs) are popular in biometric recognition tasks due to their effectiveness. Recent studies have provided evidence of the efficiency of CNNs in extracting features from images and achieving high accuracy in recognition tasks. Gupta et al. [9] demonstrated the effectiveness of CNNs in capturing complex patterns in the iris, leading to improved recognition rates. Furthermore, HD's ability to measure the similarity between binary patterns, commonly used to represent iris features, supports its use in classification. Recent research by Zhang et al. [10] highlighted the widespread use of HD in iris recognition systems due to its simplicity and efficiency in comparing binary feature vectors. By combining CNNs for feature extraction and HD for classification, the proposed hybrid approach takes advantage of the strengths of both methods to enhance the security of cloud systems. Recent experimental results have shown the effectiveness of this approach in achieving high recognition rates on various datasets, surpassing the performance of other classifiers utilized in previous studies.

Here are the paper's significant contributions, arranged pointwise.1 Theoretical Foundation:o Edge detection and segmentation techniques:⁃ Theory and Principles: The hybrid approach leverages well-established edge detection and segmentation principles to isolate and identify the iris region in eye images. The Canny edge detection identifies boundaries by detecting brightness discontinuities, whereas the Hough transform accurately delineates the circular iris area.

⁃ Application in Iris Recognition: These techniques play a crucial role in the preprocessing stages, ensuring the accurate capture of the iris region for feature extraction and biometric recognition.

o Feature extraction and classification:⁃ Hamming Distance (HD): By comparing bits at corresponding positions and counting mismatches, HD measures the similarity between binary feature vectors in iris recognition.

⁃ Convolutional Neural Networks (CNNs): CNNs, with their architecture of convolutional, pooling, and fully connected layers, enable automatic feature extraction from raw image data. Their ability to capture spatial hierarchies and local patterns is crucial for distinguishing unique iris patterns.

2. Validation Methodology:o Datasets: We validated the proposed approach using multiple publicly available iris datasets (MMU, IITD, CASIA Iris Interval V4), which cover diverse iris patterns and conditions.

o Experimental Design:⁃ We followed a rigorous experimental design by splitting datasets into training and testing sets, ensuring realistic model performance evaluation on unseen data.

⁃ Recognition Rates: To demonstrate the efficacy of the approach, we used recognition rates (accuracy) as the primary evaluation metric and made comparisons to previous research.

o We conducted systematic comparisons between the recognition rates of the proposed method and those reported in the literature, highlighting improvements and strengths.

o Statistical Validation:⁃ Cross-Validation: We performed cross-validation to ensure robustness, mitigate overfitting, and provide a comprehensive evaluation.

⁃ Statistical Significance Testing: We conducted statistical significance tests (e.g., paired t-test) to verify that recognition rate improvements were not due to random chance, thereby enhancing the credibility of the findings.

3 Performance Improvements:o The proposed approach achieved high recognition rates, particularly on the MMU V1 dataset, where CNN attained an HD accuracy of 99.50 %, surpassing the previous best HD accuracy of 95.07 %.

o Increased Classifier Accuracy: The approach enabled classifiers to achieve higher accuracy, with HD processing contributing to improved recognition rates.

4. Overall Contribution:o The mixed method that uses edge detection, segmentation, HD, CNNs, and strict validation protocols creates a strong theoretical and methodological base for using iris recognition to make cloud systems safer. The high recognition rates and their statistical validation confirm the efficacy and validity of the approach.

The paper consists of seven sections, as follows: Section 2 introduces the study's background, significance, and challenges. Section 3 presents the literature review. Section 4 introduces the proposed technique, which uses accurate iris recognition. Section 5 presents the experimental findings and analysis. Section 6 presents the limitations and future work. Finally, Section 7 concludes the work.

2 Background, significances, and challenges

Biometric authentication is a type of security technique that uses an individual's distinct biological traits to confirm their identification. This technique makes use of both physiological and behavioral characteristics, including iris patterns, fingerprints, facial recognition, voice recognition, and keystroke dynamics. Due to their intrinsic uniqueness and ability to be hard to replicate, biometrics provide greater security compared to traditional authentication techniques like passwords and PINs.

Significance of biometrics in security: Security systems with biometric authentication have various advantages including.• Biometrics offer increased security due to their difficulty in forging or stealing biological features.

• Convenience: No need for memorizing complex passwords or carrying actual tokens.

• Non-rejection: Biometric data allows for individual authentication, lowering the possibility of repudiation.

• Scalability: Biometric systems can handle huge user counts, making them ideal for cloud security and other applications.

Iris recognition as a biometric method: Iris recognition is a process that uses the iris, a colorful ring surrounding the pupil, for identification purposes. This approach enjoys high favor due to its low false match rates and great precision. The iris is a powerful biometric authentication candidate because of its rich texture and complex patterns, which do not change over a person's lifespan.

Key features of iris recognition.• Uniqueness: No two irises are identical, even among identical twins.

• Stability: Iris patterns remain stable over time, providing consistent and reliable authentication.

• Non-invasiveness: Iris recognition can be performed non-invasively using a camera, making it user-friendly.

Cloud security and the role of biometrics: The term "cloud security" refers to the practice of keeping sensitive information, programs, and services hosted on the cloud safe from hackers and other cybercriminals. Biometrics can be useful in addressing the security issues brought about by the scattered and ever-changing nature of cloud systems. Biometric authentication can improve cloud security by providing strong identity verification, decreasing the likelihood of illegal access, and strengthening access control systems.

Existing challenges in biometric authentication for cloud security: Despite its advantages, biometric authentication faces several challenges including.• Technical challenges:- Accuracy and reliability: ensuring high accuracy and low false acceptance/rejection rates, especially in varying environmental conditions.

- Scalability: efficiently managing and processing biometric data for a large number of users.

- Integration: integrating biometric systems with existing cloud infrastructure and applications.

• Security and Privacy Concerns:- Biometric Data Protection: safeguarding biometric data from theft, tampering, and misuse.

- Template Security: protecting stored biometric templates from being reverse-engineered or stolen.

- Privacy: addressing user concerns regarding the collection, storage, and use of their biometric data.

• User Acceptance:- Usability: ensuring that biometric systems are user-friendly and do not inconvenience users.

- Trust: building user trust in biometric systems through transparency and robust data protection measures.

Researchers are looking at hybrid solutions that mix biometric authentication with other forms of security to address these issues. These solutions combine the best features of many authentication systems to strengthen security, make using the system easier, and alleviate privacy issues. Here are a few examples of hybrid approaches.• Multi-factor authentication (MFA): Combining biometrics with traditional methods like passwords or smart cards.

• Multimodal Biometrics: Using multiple biometric traits (e.g., combining iris recognition with fingerprint or facial recognition) to enhance accuracy and reliability.

• Cryptographic Techniques: Encryption and secure multi-party computation are used to protect biometric data and ensure secure authentication.

3 Literatures review

In this part, the authors conducted a detailed revision of the relevant literature, which led to the discovery of some recent research on iris recognition analysis and its use in biometric authentication.

3.1 Using CNN

For C-IoT applications, the authors of [11] provided a model for iris identification based on deep CNN and chaotic encryption. They used both CASIA (the Chinese Academy of Sciences), the V4 Interval dataset, and the Phoenix dataset. The results demonstrated that the CASIA V4 and Phoenix datasets achieved the highest level of accuracy when compared to the existing approaches, up to 99.24 % and 100 %, respectively.

The authors of [12] suggested using a deep learning viewpoint to concurrently address two authentication vulnerabilities: replay and database attacks. They used the Indian Institute of Technology Delhi (IITD) and CASIA Iris Interval V4 datasets, which are two publicly accessible databases. The proposed approach performed better when compared to current modern techniques.

The authors of [13] presented an architecture based on DFCN, that is, a dense, fully convolutional network. The architecture used the Labelme software to name the eyelash areas that cover part of the iris. They utilized the IITD, CASIA Interval V4, and UBIRIS V2 databases. Precision, F1 score, recall, nice1, nice2, and accuracy error scores, as well as other metrics that measure the resilience of the proposed network, showed promising results.

The study in Ref. [14] proposes a new biometric recognition system with cancellation, leveraging the bio-convolution CNN model. The recognition rates for the databases LFW, FERET, IITD, and CASIA Iris V3 were 99.15 %, 98.35 %, 97.89 %, and 95.48 %, respectively.

The authors of [15] aimed to create an intelligent iris recognition model using deep learning and multi-objective feature selection. The authors utilized the Multimedia University (MMU) iris dataset and the IITD iris database as benchmark datasets. The suggested CNN + EH-WOA, which is an optimized method with better accuracy values than others, is 0.85 % bigger than CNN, 0.88 % bigger than NN, 0.92 % bigger than KNN, and 0.98 % bigger than Support Vector Machine (SVM).

In the field of iris identification for C-IoT applications, researchers have proposed various models and architectures based on deep learning and CNN. Researchers have evaluated these approaches using popular iris datasets like CASIA V4, Phoenix, IITD, UBIRIS V2, and the MMU iris dataset, among others. The results have shown promising accuracy levels, with some approaches achieving up to 100 % accuracy. Overall, this approach has several advantages, including.1. High accuracy: The proposed models and architectures have demonstrated high accuracy levels, reaching up to 100 % in some cases. This indicates the potential of using deep learning and CNN for iris identification in C-IoT applications.

2. Robustness: The evaluation metrics used, including precision, F1 score, recall, and other resilience metrics, indicate the resilience and robustness of the proposed approaches. This suggests that the models can perform well under different scenarios and conditions.

3. Comparative performance: Comparisons with existing techniques have shown that the proposed approaches outperform current modern techniques, indicating their superiority in terms of accuracy and efficiency.

On the other hand, this approach has several limitations/weaknesses, including.1. Dataset limitations: Despite the use of popular iris datasets like CASIA, Phoenix, IITD, and MMU for evaluation, limitations in terms of dataset size and diversity may still exist. This could affect the generalizability and real-world applicability of the proposed approaches.

2. Computational complexity: Deep learning models, especially those based on CNN, can be computationally intensive and require significant computational resources for training and inference. This might limit their deployment in resource-constrained environments or devices.

3. Model optimization: While some approaches have achieved high accuracy levels, there might still be room for further optimization in terms of model architecture, hyperparameter tuning, and feature selection. This could potentially improve the overall performance and efficiency of the proposed models.

Finally, great accuracy, resilience, and comparable performance are the benefits of iris identification in C-IoT applications with deep CNN and chaotic encryption. On the other hand, there are drawbacks, including computational complexity, dataset constraints, and the possibility of additional optimization.

3.2 Using CASIA Iris Interval V4

The authors of [16] proposed a novel approach for iris recognition based on a steerable pyramid decomposition and Gabor filter. The databases utilized were UPOL, IITD, and CASIA V4. The suggested algorithm offers 99.99 % accuracy. They have also improved the false acceptance rate (FAR), genuine acceptance rate (GAR), and equal error rate (EER).

The authors of [17] created an approach that outperforms various contemporary fusion algorithms significantly. The authors utilized EfficientNetV2, based on DropBlock, as a convolutional building block for U-Net. The experiment results of the CASIA Interval V4, UBIRIS V2, and IITD iris datasets exemplified the approach's strong adaptability and ability to deliver greater accuracy on heterogeneous databases.

In [18], the authors provided an iris segmentation method for noisy pictures using feature channel optimization. They validated the proposed architecture using the CASIA Iris Interval V4 and IITD V1 datasets. The F-measure for CASIA was 98.21 %, and the mean error measure was 0.78 %. The F-measure for IITD was approximately 97.87 %, with a mean error of 0.97 %.

The authors of [19] suggested using FRED-Net, a fully residual encoder-decoder network based on deep learning, to find the real iris area. Sending high-frequency data across residual skip connections from earlier layers or levels achieves this. They tested the suggested FRED-Net using five different iris datasets in both near-infrared (NIR) and visible light settings, as well as two general datasets for dividing road scenes. They conducted experiments using three datasets: IITD V1, CASIA Distance V4, CASIA Interval V4, and the Institute of Automation, all in an NIR light environment. The experiments in a visible light environment used two datasets: Mobile Iris Challenge Evaluation I (MICHE-I) and Noisy Iris Challenge Evaluation II (NICE-II), which selected photos using the UBIRIS V2 database. The experimental results showcased the performance of the proposed FRED-optimal Net on both the iris seven datasets and the general road segmentation scenario.

A novel iris identification feature descriptor was presented in Ref. [20]. The curvelet transform's wedge-shaped sub-bands, which give it access to the whole frequency range, were used by the descriptor to describe the iris images. To demonstrate the utility of the suggested approach, benchmark databases CASIA Iris Interval V4 and IITD were used. Results demonstrated the proposed descriptor's superior performance.

The authors of [21] suggested a method based on hybrid techniques. A simulation was run using samples from CASIA. The primary programming tool for the V4 Interval database was MATLAB. According to experimental findings, Legendre performed best in order 10, and the best result was obtained using the lower portion of the left eye, while a high recognition rate was achieved using the two sides of the right eye. The best results with the fewest errors came from combining characteristics of Legendre and LQP.

FD-UNet, the best and most efficient network model, was obtained by using the same datasets in testing and training plus U-Net combined with completely dilated convolution, as described in Ref. [22], which also offered four new possible network methods. Visible light illumination captured the iris database and NIR illumination; that is, CASIA Iris Interval V4 and ND IRIS 0405 were used to test the suggested technique, UBIRIS V2. Concerning the UBIRIS V2, ND IRIS 0405, and CASIA Iris Interval V4 datasets, F1 scores of the proposed model were 94.81 %, 97.36 %, and 96.74 %, respectively.

To address the faulty iris characteristic [23], developed a novel approach to iris identification systems based on theories of uncertainty. Interval and synthetic are two subgroups of the CASIA V4 iris photo collection. By way of Accuracy Recognition Rate (ARR), statistics exist under the area of the Receiver Operating Characteristics Curve (AUC) and EER. Experimental results demonstrated that the suggested model enhanced the recognition of the iris system.

Overall, this approach has several advantages, including.1. The proposed algorithm offers 99.99 % accuracy in iris recognition, indicating high performance [16].

2. The approach outperforms various fusion algorithms significantly, demonstrating strong adaptability and accuracy on heterogeneous databases [17].

3. The proposed architecture achieves a high F-measure and low mean error measure, indicating accurate iris segmentation [18].

4. The suggested FRED-Net network shows good performance in distinguishing true iris areas and achieves high accuracy across multiple iris datasets [19].

5. The proposed iris identification feature descriptor demonstrates superior performance compared to benchmark databases [20].

6. A method based on hybrid techniques improves recognition rate by combining characteristics of Legendre and LQP [21].

7. The proposed FD-UNet network model achieves high F1 scores on multiple iris datasets [22].

On the other hand, this approach has several limitations, including.1. Each study provides limited information on the databases used, making it challenging to evaluate the representativeness and generalizability of the results.

2. Some studies only evaluate performance on a limited set of iris datasets, which may not fully reflect real-world scenarios.

3. The computational complexity and resource requirements of the proposed algorithms are not discussed, which may affect their practicality and scalability.

4. Some studies do not compare their proposed approach with existing state-of-the-art methods, making it difficult to assess their superiority.

5. Some studies do not discuss the potential limitations or drawbacks of their proposed methods, leaving room for uncertainty in their applicability.

3.3 Using IITD

The authors of [24] suggested an improved defensive strategy to quickly identify adversarial iris samples. The Particle Swarm Optimization (PSO) technique, SVM, and Pre-trained CNN model (VGG 16) were employed. Several adversarial assaults were employed, including FGSM, iGSM, and Deepfool techniques. An experimental result on the benchmark IITD iris dataset yields exceptional results, with an average accuracy of 95.8 %. In addition to extracting and coding iris features, the authors of [25] describe an iris recognition algorithm that also incorporates the stage of feature selection. The two iris datasets, CASIA and IITD, the logistic model, and SVM were all applied. The experiment's findings demonstrated that feature selection and encoding depending upon Variable Length Black Hole Optimization (VLBHO) are superior to comparisons with a 0.21 % improvement factor.

The authors of [26] suggested a three-stage training technique. Research demonstrated that the suggested strategy enhanced segmentation performance and produced the following favorable outcomes. For CASIA T, MICHE I, UBIRIS V2, IITD, and CASIA D, outputs were 0.37 %, 0.66 %, 0.44 %, 1.03 %, and 0.41 % in Nice1 and 97.41 %, 93.21 %, 96.66 %, 98.72 %, and 94.28 % in F1.

In [27], the authors provided the protection strategy of the Block XOR-based Iris Fusion (BXBIF) template, a revolutionary cancellable technique. This non-invertible method seeks to change biometric data in a way that prevents an attacker from being capable of decrypting it. IITD database version 1.0 was utilized for the experiments, and its EER and GAR values were 0.45 and 96.7 %, respectively.

In [28], a unique dynamic route search-based preprocessing non-normalized technique for iris segmentation was investigated. The success of the suggested strategies was shown by performing benchmark tests on the IITD Iris dataset and CASIA Iris Thousand, which are open iris datasets. The authors demonstrate that when investigating identification systems of iris that use deep learning, it will be a better choice to apply iris segmentation images without normalization.

In [29], the authors go into more detail about how a one-way transformation function and an encryption function can be modified to support a biometric template with cancelation. IITD iris and CASIA Iris V4 datasets were used. The recorded accuracy value of 99.59 %, the identification accuracy value of 99.88 %, and the decreased ERR value of 0.46 % in the result analysis were encouraging. Additionally, due to its ability to identify the iris code in as short as seven ms with a true positive and true negative rate that can go up to 100 %, the computational effectiveness of the suggested method was shown.

The authors of [30] presented a solid plan. Lagrange interpolating polynomial, Hough Transform, and Viola-Jones algorithm were applied. On the iris datasets CASIA IrisV4 Distance, MMU V1, and IITD V1, and the public face dataset, this approach demonstrated increased performance. On these databases, the average accuracy rating was 97.97 %.

Overall, this approach has several advantages, including.1. The use of the Particle Swarm Optimization (PSO) technique, SVM, and Pre-trained CNN model (VGG 16) in Ref. [24] results in exceptional accuracy of 95.8 % in identifying adversarial iris samples.

2. The incorporation of feature selection and encoding using Variable Length Black Hole Optimization (VLBHO) in Ref. [25] leads to a 0.21 % improvement factor in iris recognition.

3. The three-stage training technique suggested in Ref. [26] enhances segmentation performance and produces favorable outcomes in various iris datasets.

4. The Block XOR Based Iris Fusion (BXBIF) template proposed in Ref. [27] provides a cancellable technique that protects biometric data from attackers.

5. The dynamic route search-based preprocessing non-normalized technique for iris segmentation discussed in Ref. [28] shows promising success in identification systems using deep learning.

6 [29]. offers an efficient method for biometric template cancelation, achieving high accuracy and computational effectiveness.

7 [30]. presents an approach using Lagrange interpolating polynomial, Hough Transform, and Viola-Jones algorithm that demonstrates increased performance in iris recognition on multiple datasets.

On the other hand, this approach has several limitations, including.1. The summary omits specific details about the potential limitations or drawbacks of these approaches.

2. The summary does not provide information on the computational complexity or efficiency of the proposed methods.

3. The approaches' generalizability to other datasets remains unclear, as they only mention experimental results for specific datasets.

4. The summary does not compare the proposed methods to existing state-of-the-art iris recognition techniques.

3.4 Using MMU

The authors of [31] planned, developed, and implemented three processes—segmentation, normalization, and feature encoding—as a framework that employs several subsystems to link the stages in the iris recognition system. To evaluate the framework's effectiveness, they used two sets of eye photos from an iPhone and an Android phone, along with the MMU Iris database, MICHE 1, MICHE 2, CASIA V1, and CASIA V2 iris databases, to evaluate the effectiveness of the framework. The suggested algorithm implementation to attain high performance under less-than-perfect circumstances is the focus of the current study.

In [32], the authors described a precise iris recognition system framework that uses a hybrid approach in the extraction of feature and preprocessing sections. Three tiers made up the suggested iris recognition model and used significant feature extraction. The dataset types used were UBIRIS, CASIA, and MMU. The best outcomes for carrying out the task were realized.

An efficient, quick, and reliable iris detection method that works well for eye images taken in less-than-ideal lighting conditions was suggested in Ref. [33]. Multi-scale gray level co-occurrence matrix (MSGLCM) and multi-range circle Hough transform (MRRCHT) approaches were combined. MMU V2 and CASIA V4 Distance databases were used to evaluate the offered technique. Extensive experimental results shed light on the usefulness of the proposed method for less-than-ideal iris photos.

A new system of iris identification built on methods of machine learning was presented in Ref. [34]. By utilizing oversampling before and throughout the cross-validation process, a novel recognition and authorization technique was built by utilizing the CASIA iris database of images. When IITD and MMU iris databases were used independently, the suggested generalizability technique was shown. Unification was achieved by the CASIA database, and the system operated as expected, according to the experimental study. Additionally, the computing costs are much lower compared to conventional systems, which lowers the recognition system's overall complexity and qualifies it for usage in real applications.

A trustworthy approach to iris localization maintaining speed and accuracy was suggested in Ref. [35]. Coarse iris borders were refined using the Fourier series and the Circular Hough Transform (CHT). Experimental data on three major databases, IITD V1, MMU V1, and CASIA Iris Interval, demonstrated the outperformance of the suggested approach over several iris localization techniques.

In [36], a roundness-based adjustable pupil localization method was given. Several renowned databases, including CASIA V3 light, Nice1, IITD, WVU, and MMU, were used for the experiments. The findings, based on both subjective and objective evaluation, demonstrated that the suggested strategy outperformed deep learning and advanced approaches regarding the localization capacity with few unrestricted circumstances and the time needed for running.

The summary discusses various approaches and techniques used in iris recognition systems. Here are the advantages and disadvantages of each approach mentioned:

In summary, this approach has several advantages, including.1. The segmentation, normalization, and feature encoding processes in the framework ensure accurate iris recognition. Testing on different databases enhances the framework's performance under varying conditions [31].

2. The hybrid approach in feature extraction and preprocessing improves the precision of the iris recognition system. The use of UBIRIS, CASIA, and MMU datasets maximizes the system's performance [32].

3. The suggested method is efficient, quick, and reliable in detecting iris in less-than-ideal lighting conditions. The combination of MSGLCM and MRRCHT approaches enhances its effectiveness. Testing on MMU V2 and CASIA V4 datasets validates its usefulness [33].

4. The use of machine learning and oversampling techniques improves the recognition and authorization technique of the iris identification system. The system's generalizability is demonstrated through testing on CASIA, IITD, and MMU databases. Lower computing costs make it suitable for real applications [34].

5. The suggested approach for iris localization is trustworthy, fast, and accurate. The use of Fourier series and circular hough transforms enhances the localization technique. Testing on IITD, MMU, and CASIA Iris Interval databases confirms its superiority [35].

6. The roundness-based adjustable pupil localization method outperforms deep learning and advanced approaches in terms of localization capacity and runtime. Testing on various renowned databases validates its superiority [36].

On the other hand, this approach has several limitations.1. Cons: The lack of mention of the specific algorithm implementation makes it challenging to assess its efficacy and efficiency [31].

2. There is no description of the specific techniques used in feature extraction and preprocessing, which limits our understanding of their effectiveness [32].

3. It is difficult to independently assess the effectiveness of the MSGLCM and MRRCHT approaches due to the lack of specific implementation details [33].

4. The absence of a specific machine learning algorithm makes it challenging to assess its effectiveness and performance [34].

5. The specific details of the suggested approach and its implementation are not provided, limiting our understanding of its efficiency and effectiveness [35].

6. The specific technique used for roundness-based adjustable pupil localization is not described, making it difficult to assess its effectiveness independently [36].

The proposed study addresses the limitations of using edge detection, segmentation, feature extraction, and classification methods separately by combining them into a hybrid approach that leverages the strengths of each technique. Here's how the study tackles the problem of using each method in isolation.1 Edge Detection:o Problem: When used alone, edge detection techniques like the Canny edge detector can identify the boundaries of objects within an image but may struggle to accurately isolate complex or circular shapes, such as the iris, without additional processing.

o Hybrid Solution: The proposed approach overcomes the shortcomings of edge detection when used alone to detect circular patterns by integrating edge detection with segmentation algorithms such as the Hough Transform, allowing for precise isolation and delineation of the iris region.

2 Segmentation:o Problem: Segmentation techniques can accurately outline specific regions (e.g., the iris) within an image, but they may not effectively separate these regions from noise or other irrelevant parts of the image without initial edge detection.

o Hybrid Solution: Integrating edge detection ensures that the segmentation process has a clear boundary to work with, leading to more precise delineation of the iris area and reducing errors in separating the iris from the surrounding eye regions.

3. Hamming Distance (HD):o Problem: HD is effective for measuring similarity between binary feature vectors, but when used alone, it relies heavily on the quality of the feature extraction process. Inaccurate extraction of the initial features could lead to unreliable results from the HD measure.

o Hybrid Solution: The proposed approach uses CNNs for feature extraction, ensuring high-quality features for HD comparison, thereby enhancing the accuracy and reliability of HD-based classification.

4. Convolutional Neural Networks (CNNs):o Problem: CNNs are powerful for automatic feature extraction and pattern recognition but require extensive computational resources and large amounts of data to train effectively. Additionally, they might overfit if not properly regularized or validated.

o Hybrid Solution: This method gets around these problems by adding HD as a second validation and classification step. This way, even if the CNNs' feature extraction isn't perfect, the final classification step can still use HD as a strong error-finding measure. Moreover, the use of cross-validation and multiple datasets helps prevent overfitting.

5 Overall Integration:o Problem: Using each method separately may not provide sufficient robustness and accuracy, especially in varied real-world conditions where noise, occlusions, or lighting variations could affect iris images.

o Hybrid Solution: By integrating edge detection, segmentation, CNN-based feature extraction, and HD-based classification, we create a comprehensive system that balances the strengths of each method. This results in a more reliable and accurate iris recognition system, thereby enhancing the security of cloud systems through improved biometric verification.

4 The proposed technique using accurate iris recognition

In this research study, the proposed biometric authentication technique used an accurate iris recognition system to improve cloud security. The system includes the following steps: reading the input dataset, performing the processes of edge detection, localization, and segmentation, as well as the feature extraction and classification processes. What is new in the proposed system is that it combines the two processes of segmentation and detecting edges. First, detecting edges is applied to the under-investigation image, which is then segmented to find all the edges, borders, and semicircles in the desired image. Next, the final produced image from the second step is sent to the feature extraction stage. The feature extractor and classifier operations are aided by the combination of edge detection and segmentation processes, which yields higher accuracy than when only one of these techniques is applied before classification.

4.1 Dataset features

The proposed technique is developed with the use of three different datasets, namely, MMU, IITD, and CASIA Iris Interval V4, where every image shows greyscaling in the datasets. Resolutions of 320 x 240, 320 x 240, and 640 x 480 are used for the MMU, IITD, and CASIA datasets, respectively. Fig. 1 illustrates image samples for the mentioned datasets.Fig. 1 Image samples from various training datasets: (a) Sample from IITD Dataset (b) Sample from MMU Dataset (c) Sample from Casia Dataset.

Fig. 1

4.2 Edge detection, image localization, and IRIS segmentation

Edge Detection: Detecting an edge is a procedure that is used in digital image processing to identify the image edges that are contained in the image's samples. It is processed by considering the variance in intensity in various sections of the eye image. The detection of edges is a prevalent challenge in many technological domains, including image processing and computer vision. Edge detection plays an important role in our visual abilities, allowing us to perceive depth and identify objects inside our field of view. One of the most popular edge detection methods is the Canny algorithm [37]. In the proposed approach, the Canny algorithm is employed to find the image's edges, which is essential for recognizing objects, especially circles. Considering intensity, the first derivative is used in the Canny technique for edge detection. A value of zero is assigned to places where the intensity is constant, while a value of one is assigned to places where the intensity changes quickly. The stages that make up this algorithm are listed below. At the beginning of the process, the operators use the smoothing step to blur the image after the Gaussian filter has been implemented to eliminate noise. After that, the step operator is used to extract edge information by finding the largest gradient in the image. The operator then searches for local maxima and edges them throughout the non-maximum suppression step. After that comes the twofold threshold technique with hysteresis, which is used to choose the most likely edge. At the end of the process, an image in binary levels is produced, the pixels' value being edged or unedged. All-edged pixels are often conceptualized more precisely as a collection of curves often expressed as polygons in the image domain. The results of applying edge detection on the used iris datasets images are presented in Fig. 2.Fig. 2 Implementation of Canny edge detector on three training datasets.

Fig. 2

Localization and Segmentation: The human visual system often comprises both low- and high-intensity zones. The extraction of the pupillary border is a difficult process because the gray intensity levels in these locations are extremely close to each other. It is a relief to know that the pupil and its surrounding areas have distinctive geometrical qualities and may be distinguished from one another. After the phase of Canny edge detection is complete, a coarse iris region undergoes localization and segmentation at the same time. Image segmentation employs similarities between parts of an image to separate an image into distinct regions. It is typically employed in the context of image and boundary detection.

As a method of data segmentation, the Hough Circle Transform tool is utilized in this particular study. Two Circular Hough Transforms are applied. The first time, the whole eye is utilized to locate the iris/sclera boundary. The second time, the iris/pupil border is determined using the iris area. In Hough space, the Circular Hough Transform produces at each detected edge a circle with unique radii. Hough space's highest point has coordinates that correspond to the center of the circle defined by the points of edge and radius. It is easier to express the circles by parameter space than lines because parameters for circles may be converted into the new space instantly. Canny's approach is utilized, as mentioned in the preceding section, to define the borders of the image as well as to help in the identification of features inside the image, most commonly circles, by making use of the Hough Transform. To successfully detect edges, it was necessary to consider the intensity differences that exist between various parts of an image. After the edges are determined, figures comprising a portion, or the entire circle are recognized to locate the iris. However, due to factors such as eyelids, eyelashes, or the use of spectacles, only a portion of this circumference is visible in the photograph. The ability to recognize edges is essential when determining circumference for use in iris recognition. The results of localization and segmentation procedures performed on the mentioned three datasets are presented in Table 1.Table 1 Implementing the localization and segmentation of the training Datasets.

Table 1Training Dataset	Source Image	Localization	Segmentation	
CASIA Iris V4	Image 1	Image 2	Image 3	
IITD V1	Image 4	Image 5	Image 6	
MMU V1	Image 7	Image 8	Image 9	

4.3 Feature extraction and classification

The procedure for determining which features to extract from the pattern of the iris is an extremely significant part of the overall process of iris recognition. The level of recognition achieved when comparing two iris templates is most strongly influenced by the feature extraction process. In this research, we employ CNN to analyze a segmented iris image for useful information. A matching metric is required to conduct a comparison of the degree of similarity between two iris templates. This matching metric tells us if two templates were created by the same individual or two separate people. In this instance, the feature extraction and classification phases are carried out using either CNN or HD. After edge detection and segmentation, the iris image is fed into the CNN or HD system as the input.

4.3.1 The deep CNN model

The CNN model has been built using two distinct datasets: one for training and one for evaluation. Each subfolder of the training dataset contains images of the iris of an individual, and the CNN model analyzes these images for unique features. This dataset is typically significantly more extensive than the set of data that is used for testing. More importantly, the larger the percent used in training, the better the accuracy of the results generated by using the trained dataset. The training dataset is utilized by the machine to discover various features, and layers help in the feature extraction process. In this context, the term “layer” refers to a specific method that modifies the appearance of the image by altering its size, shape, color, or other features by making use of pixels that are far smaller than the actual image. Images in the same category are given numerous layers of enhancement before being saved in the CNN. The training dataset is used to determine the chance of certain features being repeated, which is then used to inform the stored features. In this work, a CNN architecture of the most basic kind is described. Convolutional layers informed by empirical research have been used in its construction. The development of a solid CNN architecture is required if one is to accomplish an iris representation that is both efficient and effective. Here, the created network is robust and has several convolutional layers, and it is fed iris greyscale images that have a resolution of 320 x 240 pixels. The process of batch normalizing occurs after each convolutional layer in a CNN, beginning with conv1 and continuing through conv5. Typically, a neural network will do pooling once every two convolutional layers, for all four layers in the CNN model. The construction of the network is initially designed by piling on successive convolutional layers in a stacked fashion. In addition, the first and second operations of pooling are carried out after the convolution computation layer, and the third operation called (pool3) is carried out as soon as the fifth layer in the CNN (conv5) is completed. In most cases, extremely minute kernels of convolution having the size of 3 x 8 (with a stride equal to 2, padding with ‘same’ values) are utilized in the modeling process. The max-pooling operation is carried out using a stride 2 over a 2 x 2-pixel frame everywhere in the network. Each results in the neuron (output) being attached to all input neurons at the top triple layers, to have full connectivity between them. The output of the Softmax classifier is passed on to the final layer that is completely linked. The learning rate starts at 0.01, which is the default setting in this study, regardless of the dataset, and then is factored by a tenth whenever there is no further enhancement in the validation error rate. During the training process, a total of 15 epochs are used and then shuffled. After testing more than 15 epochs, we determined that the precision was not up to par. The Rectified Linear Unit (ReLU) activation function is employed across all hidden layers in deep neural networks. The stochastic gradient descent (SGD) algorithm is carried out throughout the training phase, and the network gradients are calculated via the back-propagation technique.

4.3.2 The HD metric

The feature extraction stage and classification approach can use the HD metric to determine if two iris samples are from the same iris or distinct iris types. The HD value is calculated by comparing the Y bits to the X bits to determine how many distinct bits there are relative to the whole bits in the template. In this study, ⊕ is used to represent the logic XOR operation, while N is used to represent the feature-vector length. The symbols Xj and Yj are examples of the compared binary codes applied to the two distinct templates from image features codes, respectively:(1) HD=1N∑j=1N(Xj⨁Yj)

After the edges of the image have been identified and separated, the noise mask and the features template that corresponds to it are then generated. The HD can be calculated by comparing this final template to all of the other final templates in the experiment. The HD between iris templates made from different irises should be greater than or equal to 0.35, while the HD with identical irises is extremely close to 0. Since the two-bit pattern values are entirely random, therefore the HD for the identical irises must be extremely close to zero and the HD for different irises must be larger than or equal to 0. Based on our model, we anticipate that the HD will range from 0.14 to 0.35 for in-eye iris images and 0.36 to 0.56 for out-of-eye iris photos. The trials detailed in the subsequent section serve as the basis for the HD range used.

5 Experimental findings and analysis

In addition to CNN and HD, the proposed system presents a powerful hybrid model integrating edge detection with segmentation to extract classification and features. The suggested approach is tested using MMU, IITD, and CASIA Iris Interval V4 datasets. Using HD on CASIA, CNN on IITD, and MMU, the biometric proposed model demonstrated gratifying high accuracies for person recognition of 99.50 %, 97.18 %, and 95.07 %, respectively. The obtained accuracy results demonstrated the classifier's superiority over others utilized in the published literature.

In Table 2, you can see how well the proposed method using CNN for biometric authentication worked on the IITD and MMU datasets compared with previous work. Table 3 illustrates the three training datasets used for localization and segmentation. As shown in Table 4, the IITD dataset achieved classification accuracy of 97.18 % and 99.50 % using the proposed classifiers, which is better than the performance of the previous approaches. Table 5 shows that, in contrast to previous studies, the proposed CNN classifier achieved a testing success rate of 99.50 % when applied to the MMU dataset. Table 2, Table 3, Table 4, Table 5 compare the methodology, datasets, and accuracy of the associated work. Fig. 3, Fig. 4 illustrate the accuracy of IITD and MMU using the proposed approach.Table 2 Researchers used CNN to address iris recognition problems.

Table 2Ref.	Technique	Accuracy	
[11]	For C-IoT applications, chaotic encryption, and deep CNN were used. The proposed technique offered a short training period, which is an important factor in crucial IoT-based applications like telemedicine.	Accuracy was obtained with CASIA V4 at 99.24 % and Phoenix datasets at up to 100 %.	
[12]	The authors suggested using deep learning to concurrently address two authentication threats, the replay attack, and the database attack.	Accuracy for the CASIA Iris Interval V4 database was 99.72 % and for the IITD database was 99.39 %.	
[13]	First, the authors presented an architecture based on densely convolutional networks (DFCN). Secondly, they used the Labelme software program for tagging the eyelashes that may occlude part of the iris.	Tests using UBIRIS V2, CASIA Interval V4, and IITD showed good results. The suggested iris segmentation network was shown to be superior concerning several criteria.	
[14]	The authors suggested a new biometric authentication system depending on deep CNN and the bio-convolution model.	The following rates for recognition were reached: 99.15 % LFW, 98.35 % FERET, 97.89 IITD, and 95.48 % CASIA Iris V3 databases.	
[15]	The authors wanted to combine deep learning with multi-objective feature selection to create an intelligent iris identification algorithm.	To evaluate the suggested model's success across various network configurations for precise iris detection, experiments were carried out on benchmark datasets. With the learning rate set at 75 % for the iris dataset MMU, suggested improved CNN + EH-WOA showed greater accuracy values using experimental analysis, scoring better and greater results than CNN, KNN, SVM, and NN.	
Proposed method	The study proposed a powerful hybrid approach that combines edge detection and segmentation, in addition to CNN and HD for classification and feature extraction.	Using CNN on IITD, and MMU, the proposed technique demonstrated gratifying biometric identification accuracies of 97.18 %, and 99.50 %, respectively.	

Table 3 Implementation of localization and segmentation on three training datasets.

Table 3Ref.	Technique	Accuracy	
[16]	The authors proposed a new method for steerable pyramid decomposition based on Gabor filters for iris recognition.	Databases from UPOL, CASIA V4, and IITD were utilized to validate the proposed approach. The suggested algorithm offered 99.99 % accuracy. Additionally, FAR, EER, and GAR were enhanced.	
[18]	Noisy photos were optimized by a featuring channel.	The mean error rate for the CASIA V4 period was 78 hundredth %, the F-measure equal to 98.21 %, and the mean error rate with the IITD V1 was 97 hundredth %, the F-measure equal to 97.87 %. The suggested method was evaluated better than the most recent techniques in noisy environments.	
[19]	The genuine iris area could be identified using a deep learning-based FRED-Net, suggested to do so while involving the transmission of high-frequency data over a residual skip link from earlier layers.	Five separate iris datasets in visible and NIR light settings, as well as two generic road scene segmentation datasets, were manipulated to test the precision of the presented FRED-Net. The experimental findings demonstrated the proposed FRED-Net's best performance on general road scene segmentation and seven iris datasets.	
[20]	The study displayed a brand-new iris recognition feature descriptor. The curvelet transform's wedge-shaped subbands, which give it access to the whole frequency range, were utilized by descriptors to characterize the iris photos.	To demonstrate the utility of the suggested approach, benchmark databases CASIA Iris Interval V4 and IITD were used. Performance indicators like DI, AUC, FAR, and EER were used to evaluate the proposed approach. Results amply demonstrated the proposed descriptor's superior performance.	
[21]	A method based on hybrid techniques was suggested by the authors. The Legendre moment, Circular Hough Transform, and Local Quantized, among other methods, were applied.	CASIA V4 Interval database samples were utilized in simulation; MATLAB served as the primary programming language. The best results with the fewest errors came from combining the characteristics of Legendre and LQP.	
[22]	The authors suggested four new workable network designs, with FD-UNet, the best and most efficient network model obtained by using the same datasets in testing and training plus U-Net combined with completely dilated convolution.	The suggested model's F1 scores on the UBIRIS V2, CASIA Iris Interval V4, and ND IRIS 0405 datasets were recorded by values of 94.81 %, 97.36 %, and 96.74 %, respectively.	
[23]	To address one defective iris characteristic, the authors presented a novel way for iris identification systems based on uncertainty theories.	On the Interval and Synthetic CASIA V4 iris' subsets picture collection, various comparative tests were carried out. By way of EER, ARR, and AUC statistics, experimental results demonstrated that the suggested model enhanced the recognition system of the iris.	
Proposed
Method	The research introduced a novel hybrid approach that combines edge detection and segmentation with convolutional neural networks and HD for classification and feature extraction.	Using HD on CASIA, CNN on IITD, and MMU, the suggested biometric approach demonstrated gratifying accuracies of 95.07 %, 97.18 %, and 99.50 %, respectively, for recognition.	

Table 4 Researchers used IITD to address iris recognition problems.

Table 4Ref.	Technique	Accuracy	
[24]	To effectively identify the adversary iris examples, an optimum defense strategy was suggested. VGG 16 with pre-trained parameters, SVM, and PSO-SVM were employed.	The suggested system was put to the test using a variety of adversarial assaults, including FGSM, iGSM, and Deepfool techniques. The result of experimentation on the benchmark IITD iris dataset yielded exceptional results, with a typical precision of 95.8 %.	
[25]	The authors offered a novel iris recognition method that included a process of feature selection in addition to iris feature coding and extraction. VLBHO, logistic model, and SVM were employed.	Results were assessed using CASIA and IITD. They showed that VLBHO outperformed the benchmarks by 0.21 %.	
[26]	Three stages of strategy training were suggested by the authors.	The segmentation performance was enhanced, according to experiments, and the proposed technique produced more promising results. For CASIA T, MICHE I, UBIRIS V2, IITD, and CASIA D, outputs were 0.37 %, 0.66 %, 0.44 %, 1.03 %, and 0.41 % in Nice1, and 97.41 %, 93.21 %, 96.66 %, 98.72 %, and 94.28 % in F1.	
[27]	The template of the Block XOR-Based Iris Fusion (BXBIF) protection strategy was a brand-new cancellable technology.	The IITD database version 1.0 was employed in the experiments, and its EER and GAR values were 0.45 and 96.7 %, respectively.	
[28]	The authors looked into a brand-new iris segmentation dynamic path search-based nonnormalized preprocessing technique.	Results showed that suggested strategies were effective through testing benchmarks on CASIA and IITD iris datasets. The authors demonstrated that the option to employ iris segmentation photos with no normalization was suitable when investigating iris identification systems depending on deep learning.	
[29]	Using a one-way transformation operation and an encryption mechanism, the indicated research's goal was to construct a cancelable transformed iris biometrical template.	The IITD and CASIA Iris V4 datasets were utilized in the experiment. The accuracy was 99.59 %, the identification rate was 99.88 %, and the ERR of 0.46 % in the result analysis was encouraging.	
[30]	The authors presented a reliable method that, among other things, segmented eyes using geometrical data and the algorithm of Viola-Jones from the human face.	This approach demonstrated an increased performance. Using MMU V1, IITD V1, and CASIA Iris V4 Distance datasets, it achieved an accuracy rating of 97.97 % on average.	
Proposed
Method	This research introduced a hybrid approach that combines edge detection and segmentation with convolutional neural networks and HD for classification and feature extraction.	Using CNN on IITD, and MMU, the presented biometric approach demonstrated gratifying recognition rates of 97.18 %, and 99.50 % respectively.	

Table 5 Researchers used MMU to address iris recognition problems.

Table 5Ref.	Technique	Accuracy	
[31]	Each stage of the proposed study's framework, segmentation, normalization, and feature encoding had relations to other iris system recognition. The planning and implementation were done using a variety of subsystems.	Two eye photo groups, including those taken with an iPhone and an Android phone, the MICHE I, MICHE II, CASIA V1, MMU, and CASIA V2 Iris databases, were used to enhance the proposed framework performance.	
[32]	Concerning feature extraction and preprocess sections of this research, a precise iris recognition framework was presented.	The suggested model was applied to three different dataset types, including UBIRIS, CASIA, and MMU, and achieved the best performance for carrying out activities. For the best possible outcomes, the error rate was examined together with various statistical metrics.	
[33]	The authors provided an efficient, quick, and reliable iris detection method that worked well for ocular photos captured in less-than-ideal conditions. The MRRCHT and MSGLCM approaches were employed.	The authors used the CASIA V4 Distance Iris database to create and compare four databases. Iris recognition accuracy results were evaluated at 87.6761 %.	
[34]	A brand-new iris identification system built on machine learning techniques was presented. The interconnectedness of biometric systems and stylometry served as the driving force of this research.	The proposed system operated as expected, according to the experimental study. Additionally, the computing costs were much lower compared to conventional systems, which lowered the recognition system's overall complexity and qualified it for usage in real applications.	
[35]	The authors suggested a reliable iris localization method that maintained accuracy and speed.	Experiments showed an accuracy of 99.53 % for IITD V1.0, 99.30 % for CASIA Iris V3 Interval, and 99.34 % for MMU V1.0.	
[36]	The roundness criteria-based adaptive pupil localization approach was presented.	After the suggested algorithm's performance was evaluated, it became clear that it guaranteed quick and reliable pupil localization even when corneal reflection, weak contrast, spectacles, or eyelashes obstruction were present: accuracy was 84.761 % for Nice1, 95.263 % for CASIA V3 light, 90.244 % for MMU, 93.618 % for WVU, and 94.316 % for IITD.	
Proposed
Method	In this study, we combined edge detection and segmentation with convolutional neural networks and high-dimensional feature extraction to create an accurate and powerful hybrid approach.	Using HD on CASIA, and CNN on MMU, the presented technique demonstrated gratifying recognition accuracies of 95.07 %, and 99.50 %, respectively.	

Fig. 3 IITD dataset accuracy of 97.18 % using the proposed approach.

Fig. 3

Fig. 4 MMU dataset accuracy of 99.50 % using the proposed approach.

Fig. 4

6 Limitation and future work

The proposed approach may encounter the following potential limitations.1. Dataset Specific Performance: The method demonstrates high recognition rates on specific datasets (MMU, IITD, CASIA), but it does not address its performance on other, potentially more diverse, or real-world datasets.

2. Overfitting: High performance on the datasets used could indicate overfitting, as the model may not generalize well to new, unseen data.

3. Resource-Heavy: Putting together edge detection, segmentation, feature extraction, and CNNs can be very hard on the computer, needing a lot of memory and processing power, which might not be possible for real-time uses.

4. Scalability Issues: The method might struggle to scale efficiently with increasing data size and complexity, particularly in large-scale cloud systems.

5. Biometric Spoofing: Despite its general security, iris recognition remains susceptible to spoofing attacks, such as the use of high-quality images or artificial eyes, which remain unaddressed.

6. CNN Interpretability: People frequently perceive CNNs as "black boxes" with limited interpretability, which makes it challenging to fully comprehend and trust their decision-making process.

For future work, a suggestion is to integrate cloud security features such as data encryption, secure authentication protocols, and robust access controls into the biometric system. By enhancing the authentication process, we can better secure sensitive data and prevent unauthorized access. This approach can help to improve the overall security of cloud systems and better protect user information. Also, we are working on providing a more detailed analysis and comparison with existing methods in the future to better highlight the strengths of our proposed approach.

Our future study will also focus on the utilization of the most recent methods, including.1. An advanced multi-modal ocular identification system that improves recognition accuracy by integrating spatial feature reconstruction with unsupervised image quality assessment [38].

2. Looking into how convolutional neural networks (CNNs) can be used to improve identification rates and show that they have potential for safe biometric uses by getting detailed information from iris images [39].

3. Enhancing recognition performance by implementing a unique data-knowledge-driven learning strategy for iris recognition [40]. This strategy integrates data-driven techniques with knowledge-based methodologies.

4. Enhancing cloud security by utilizing AuthPrivacyChain and a hybrid encryption strategy, allows for more secure authentication processes and better scalability [41].

5. DL-based hybrid cryptography develops a mutual authentication mechanism for cloud data security. By integrating DL methods with hybrid cryptographic algorithms, the approach guarantees safe communication between cloud users and services [42].

6. A secure authentication approach utilizing a multimodal biometric recognition system based on DL. The system achieves high authentication security and accuracy by combining many biometric modalities (e.g., iris, face, and fingerprint) and using deep learning algorithms [43].

7 Conclusion

The theoretical foundation of our approach was centered around two main techniques: (1) Edge detection and segmentation techniques: (a) Theory and principles: The proposed hybrid approach uses well-established principles of edge detection and segmentation to isolate and identify the iris region from eye images. Edge detection algorithms, such as the Canny edge detector, rely on mathematical constructs that identify boundaries within images by detecting brightness discontinuities. We employ segmentation techniques, such as the Hough Transform for circular shapes, to accurately delineate the iris area from the rest of the eye. (b) Application in iris recognition: These techniques play a crucial role in preprocessing steps, guaranteeing the accurate capture of the iris region, which is essential for reliable feature extraction and subsequent biometric recognition. (2) Feature extraction and classification: Various domains, particularly in error detection and correction, use HD as a measure. In the context of iris recognition, it quantifies the similarity between binary feature vectors extracted from iris images. The theoretical underpinning is based on comparing bits at corresponding positions and counting the number of mismatches. CNNs are a class of deep learning models specifically designed for image analysis tasks. The architecture of CNNs, comprising convolutional layers, pooling layers, and fully connected layers, enables automatic feature extraction from raw image data. The theoretical foundation of CNNs lies in their ability to capture spatial hierarchies and local patterns within images, which is crucial for distinguishing unique patterns in iris recognition.

Our proposed approach includes the following validation methodology: (1) Datasets: We used multiple publicly available and widely accepted iris datasets, including MMU, IITD, and CASIA Iris Interval V4, to validate our proposed approach. We tested our method against various conditions and variations in iris patterns using these datasets, which provide a diverse range of iris images. (2) Experimental design: (a) Training and testing split: We followed a rigorous experimental design by dividing each dataset into separate training and testing sets. This split guarantees the evaluation of the model on unseen data, thereby offering a realistic assessment of its performance. (b) We used recognition rates (accuracy) as the primary metric to evaluate our method. We calculated the recognition rates for each dataset and compared them with those of previous research to demonstrate the efficacy of our approach. (3) Comparison with previous research: We systematically compared the recognition rates achieved by our proposed method with those reported in the literature. This comparative analysis highlights the improvements and strengths of our approach. (4) Statistical Validation: (a) Cross-validation: To ensure robustness, we performed cross-validation where applicable. This technique entails partitioning the dataset into multiple folds and iteratively training and testing the model on different folds. Cross-validation helps mitigate overfitting and provides a more comprehensive evaluation. (b) Statistical significance testing: We employed statistical significance tests, such as the paired t-test, to verify that the improvements in recognition rates were not due to random chance. This rigorous statistical analysis increases the credibility of our findings.

In conclusion, our hybrid approach, combining edge detection, segmentation, feature extraction using HD and CNN, and subsequent validation using well-defined datasets and rigorous experimental protocols, establishes a solid theoretical and methodological foundation. The high recognition rates achieved, and their statistical validation confirm the efficacy and validity of our proposed approach for enhancing the security of cloud systems through iris recognition.

Finally, our approach helped the classifiers achieve a higher degree of accuracy than they had before. Because HD processed the iris image, the accuracy was higher than in previous trials. On the MMU V1 dataset, CNN successfully attained an HD accuracy of 99.50 %, surpassing the best HD accuracy of 95.07 %. As a result, the proposed approach has a better recognition rate in terms of accuracy.

Ethical approval

Not applicable.

8 Data availability statement

Data will be made available on request.

CRediT authorship contribution statement

Hosam El-Sofany: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Belgacem Bouallegue: Writing – review & editing, Validation, Methodology. Yasser M. Abd El-Latif: Writing – review & editing, Validation, Methodology.

Declaration of competing interest

The authors declare that they have no conflicts of interest to report regarding the present study.

Acknowledgment

The authors extend their appreciation to the Deanship of Research and Graduate Studies at 10.13039/501100007446 King Khalid University for funding this work through small group research under grant number (RGP1/129/45 ).
==== Refs
References

1 Milad S. Najafzadeh S. Authentication techniques in cloud computing: a review Int. J. Adv. Res. Comput. Sci. Software Eng. 1 2017 95 99
2 Tiwari S. Chaurasia J.N. Chaurasia V.S. A review of advancements in biometric systems Int. J. Innov. Res. Adv. Eng. (IJIRAE) 2 1 2015 187 204
3 Jagadeesh N. Patil Chandrasekhar M. Iris recognition system development using MATLAB International Conference on Computing Methodologies and Communication (ICCMC) 2017 IEEE 348 353
4 Kumar V. Ahmad M. Kumari A. Kumari S. Khan M.K. SEBAP: a secure and efficient biometric‐assisted authentication protocol using ECC for vehicular cloud computing Int. J. Commun. Syst. 34 2019
5 Mala R. Kumar C.J. Providing authentication by using biometric multimodal framework for cloud computing, TELKOMNIKA Indones J. Electr. Eng. 15 3 2015
6 Mansour A. Sadik M. Sabir E. Multi-factor authentication based on multimodal biometrics (MFA-MB) for Cloud Computing 2015 IEEE/ACS 12th International Conference of Computer Systems and Applications (AICCSA) 2015 1 4
7 Kavitha R. Valarmathi A. A security hybrid mechanism using fuzzy with adaptive ElGamal approach in cloud Concurrency Comput. Pract. Ex. 34 2022
8 Sen P. Prasad R. Saurabh P. A new approach for cloud security using hybrid querying system over cloud scenario Adv. Intell. Syst. Comput. 904 2019 367 376
9 Gupta R. Agrawal R.K. Mittal A. Biometric recognition: a state-of-the-art review, Journal of King Saud University-Computer and, Information Sciences 36 5 2018
10 Zhang Z. Zhong K. Wang L. Chen Y. Iris recognition based on local features learning Inf. Sci. 47 6 2016
11 Shalaby A.S. Gad R. Hemdan E.E.D. An efficient CNN based encrypted Iris recognition approach in cognitive-IoT system Multimed. Tool. Appl. 80 2021 26273 26296 10.1007/s11042-021-10932-x
12 Gupta Richa Sehgal HsIrisNet Priti Histogram based Iris recognition to allay replay and template attack using deep learning perspective Pattern Recogn. Image Anal. 30 2020 786 794 10.1134/S105466182004015X
13 Chen Y. Wang W. Zeng Z. Wang Y. An adaptive CNNs Technology for robust Iris segmentation IEEE Access 7 2019 64517 64532 10.1109/ACCESS.2019.2917153
14 Abdellatef E. Soliman R.F. Omran E.M. Cancelable face and iris recognition system based on deep learning Opt. Quant. Electron. 54 2022 702 10.1007/s11082-022-03770-0
15 Babu G. Khayum P.A. Elephant herding with whale optimization enabled ORB features and CNN for Iris recognition Multimed. Tool. Appl. 81 2022 5761 5794 10.1007/s11042-021-11746-7
16 Mohamed Radouane Zouggari Nadia Idrissi Amraoui Amine Mounir Amraoui Fusion of Gabor filter and steerable pyramid to improve iris recognition system Int. J. Artif. Intell. 11 4 2022 1460 1468 10.11591/ijai.v11.i4.pp1460-1468
17 Huo Guang Lin Dawei Yuan Meng Yang Zhiqiang Niu Yueqi Heterogeneous iris segmentation method based on modified U-Net J. Electron. Imag. 30 6 25 November 2021 063015 10.1117/1.JEI.30.6.063015
18 Hao K. Feng G. Ren Y. Iris segmentation using feature channel optimization for noisy environments Cogn Comput 12 2020 1205 1216 10.1007/s12559-020-09759-9
19 Arsalan Muhammad Kim Dong Seop Lee Min Beom Owais Muhammad Park Kang Ryoung FRED-Net: fully residual encoder–decoder network for accurate iris segmentation Expert Syst. Appl. 122 2019 217 241 10.1016/j.eswa.2019.01.010 ISSN 0957-4174
20 Vyas Ritesh Kanumuri Tirupathiraju Sheoran Gyanendra Dubey Pawan Efficient iris recognition through curvelet transform and polynomial fitting Optik 185 2019 859 867 10.1016/j.ijleo.2019.04.015 ISSN 0030-4026
21 Hashim A.N. Al-Hashimi B.M. Human Iris recognition based on hybrid technique J. Comput. Sci. 15 12 2019 1734 1745 10.3844/jcssp.2019.1734.1745
22 Zhang W. Lu X. Gu Y. Liu Y. Meng X. Li J. A robust Iris segmentation scheme based on improved U-Net IEEE Access 7 2019 85082 85089 10.1109/ACCESS.2019.2924464
23 Bellaaj M. Khanfir Kallel I. Sellami D. Probability-possibility theories based iris biometric recognition system ELCVIA - Electron. Lett. Comput. Vis. Image Anal. 18 1 2019 20 37
24 Meenakshi K. Maragatham G. An optimized defensive technique to recognize adversarial iris images using curvelet transform Intelligent Automation & Soft Computing 35 1 2023 627 643
25 Saraf T.O.Q. Fuad N. Taujuddin N.S.A.M. Feature encoding and selection for Iris recognition based on variable-length black Hole optimization Computers 11 2022 140 10.3390/computers11090140
26 Zhou Z. Liu Y. Zhu X. Liu S. Zhang S. Li Y. Supervised contrastive learning and intra-dataset adversarial adaptation for Iris segmentation Entropy 24 2022 1276 10.3390/e24091276 36141162
27 Rafiq S. Selwal A. Block - XOR based cancellable template protection scheme for multi-instance iris biometric system Multimed. Tool. Appl. 81 2022 23217 23235 10.1007/s11042-022-12655-z
28 Chen Yifeng Wu Cheng Wang Yiming Whether normalized or not? Towards more robust iris recognition using dynamic programming Image Vis Comput. 107 2021 104112 10.1016/j.imavis.2021.104112 ISSN 0262-8856
29 Rajasekar V. Premalatha J. Sathya K. Cancelable Iris template for secure authentication based on random projection and double random phase encoding Peer-to-Peer Netw. Appl. 14 2021 747 762 10.1007/s12083-020-01046-6
30 Jan F. Alrashed S. Min-Allah N. Iris segmentation for non-ideal Iris biometric systems Multimed. Tool. Appl. 2021 10.1007/s11042-021-11075-9
31 Tawfik Mohammed, Reend & Kaur, Harleen & Alankar, Bhavya & Chauhan, Ritu. (2022). Recognition of human Iris for biometric identification using Daugman's method. IET Biom. 11. n/a-n/a. 10.1049/bme2.12074.
32 Ulah Arif Salam Abdu el Raoui Hanane Sebai Dorsaf Rafie Mahnaz Towards more accurate iris recognition system by using hybrid approach for feature extraction along with classifier Int. J. Reconfigurable Embed. Syst. 11 2022 59 70 10.11591/ijres.v11.i1.pp59-70
33 Nsaif Ahmed Khudhur Ali Sawal Hamid Md Nseaf Asama Kuder Jassim Khider Nassif Al-Qaraghuli Ammar Sulaiman Riza Robust and Swift Iris Recognition at based on novel pupil segmentation Journal of King Saud University - Computer and Information Sciences 2022 10.1016/j.jksuci.2022.09.002 ISSN 1319-1578
34 Adamović Saša Miškovic Vladislav Maček Nemanja Milosavljević Milan Šarac Marko Saračević Muzafer Gnjatović Milan An efficient novel approach for iris recognition based on stylometric features and machine learning techniques Future Generat. Comput. Syst. 107 2020 144 157 10.1016/j.future.2020.01.056 ISSN 0167-739X
35 Jan F. Min-Allah N. Agha S. A robust iris localization scheme for the iris recognition Multimed. Tool. Appl. 80 2021 4579 4605 10.1007/s11042-020-09814-5
36 Gautam G. Mukhopadhyay S. An adaptive localization of pupil degraded by eyelash occlusion and poor contrast Multimed. Tool. Appl. 78 2019 6655 6677 10.1007/s11042-018-6371-0
37 Song R. Zhang Z. Liu H. Edge connection based Canny edge detection algorithm Pattern Recogn. Image Anal. 27 4 2017 740 747
38 Yan Z. Wang Y. Zhang K. Sun Z. Boosting multi-modal ocular recognition via spatial feature reconstruction and unsupervised image quality estimation Machine Intelligence Researc 21 2024 197 214
39 Sallam A.A. Amery H.A. Saeed A.Y.A. Iris recognition system using deep learning techniques Int. J. Biometrics. 15 6 2023
40 Liu S. Liu Y. Zhu X. Zhang S. Data-knowledge driven: a new learning strategy for iris recognition Multimed. Tool. Appl. 83 9 2024 1 31
41 Ananthakrishna V. Yadav C.S. Innovations in cloud security: enhanced hybrid encryption approach with AuthPrivacyChain for enhanced scalability, Nanotechnology Perceptions 20 52 2024 560 575
42 Mohd A.A. Kummarikunta S. Design of mutual authentication method for deep learning based hybrid cryptography to secure data in cloud computing Journal of Safety & Security Engineering 13 2023 893 902
43 Rajasekar V. Saracevic M. Hassaballah M. Efficient multimodal biometric recognition for secure authentication based on deep learning approach Int. J. Artif. Intell. Tool. 2023 2023 World Scientific
