A Comprehensive Review of Deep Learning- Driven Facial Recognition Frameworks for Intelligent and Fully Automated Attendance Monitoring Systems

A Comprehensive Review of Deep Learning- Driven Facial Recognition Frameworks for Intelligent and Fully Automated Attendance Monitoring Systems | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Comprehensive Review of Deep Learning- Driven Facial Recognition Frameworks for Intelligent and Fully Automated Attendance Monitoring Systems Raj Ranjan, Sureshwati MCA, Manish Kumar, Rahul Gandhi, Pawan Kumar, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9504761/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The challenge for schools and organizations in need of an intelligent, automated, and real-time attendance monitoring system has become complex. In this survey, we present a survey of facial recognition frameworks based on deep learning and addressed the architectural patterns of CNNs (CNN), ResNets, FaceNet, ArcFace, CosFace, and Vision Transformers as well as face detection pipelines based on MTCNN, YOLOv5, and YOLOv8. We discuss the steps of a face detection pipeline—face detection, preprocessing, feature extraction, matching, and database integration, and practical deployment scenarios. Our study demonstrates that deep learning models such as ArcFace and transformer models provide high accuracy for facial identity verification. Transfer learning techniques have effectively removed the need for large labeled institutional datasets while light versions of FaceNet and MobileNet-based models allow real- time processing on edge devices. Further, multi-face simultaneous recognition models with anti-spoofing modules provide more reliability for large- scale attendance applications. But, several obstacles are preventing widespread adoption. Performance under occlusion and varying illumination, demographic bias, privacy and regulatory concerns, and computational limitations still limit the adoption of attendance monitoring. Future research should focus on robust cross-environment models, fairness- aware training, edge-optimized architectures, and privacy- preserving models to ensure responsible and scalable attendance monitoring. Facial recognition deep learning attendance monitoring ArcFace FaceNet MTCNN YOLOv8 convolutional neural networks transformer models biometric systems edge deployment Figures Figure 1 Figure 2 Figure 3 I. INTRODUCTION The advent of artificial intelligence and computer. vision technologies has revolutionized the ways. businesses check their identity and access control. One of the attendance is by far the largest example of this new era. monitoring. The facial recognition technology (FRT) enables. organizations to determine people in real-time without the requirement. proximity cards, or manually entering names into a roll call box [ 1 ]. FRT is much more safe than finger or iris scanning. Face recognition can be used in large-scale environments such as universities, office buildings, and health canter’s. During the past decade, deep learning has revolutionized facial recognition. CNNs, ResNets, FaceNet, ArcFace, and Vision Transformers consistently achieve near-human or superior accuracy on standard benchmark datasets [ 2 ]. From handcrafted features like Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) to deep learned embeddings, we have been able to generalize across lighting, pose, and occlusion. Meanwhile, Multi-task Cascaded Convolutional Networks (MTCNN) and YOLO detectors have enabled simultaneous multi-face detection in real time video, which is necessary to accurately track attendance in crowded classrooms or meetings [ 3 ]. Despite these advances, deep learning-based attendance systems can still be difficult to deploy on an institutional scale. For instance, facial recognition with partial face occlusion due to masks or glasses, varying lighting, morphology, and low resolution input data can reduce performance dramatically [ 4 ]. Privacy issues with biometric data storage, protection against spoofing, and requirements for compliance with regulations such as GDPR and EU AI Act (2024) continue to hinder large- scale application [ 5 ]. Review of the literature reveals many independent applications of deep learning, but few recent and current review studies that are looking at the architectures, face detection pipelines, benchmark datasets, real world deployment constraints, ethical considerations, and performance trade-offs in attendance monitoring. This paper attempts to fill this gap by providing a systematic review of deep learning-based facial recognition applied to attendance monitoring. It covers the entire process of recognition from face detection and preprocessing to feature extraction, matching, and system integration. II. LITERATURE REVIEW AND RELATED WORK A. Traditional and Classical Approaches to Attendance Monitoring Initial attendance systems were based on proximity technologies such as RFID cards, barcode scanning, and PIN- based entry. These are functional, but allow for proxy attendance (when someone marks one’s attendance for another) and require physical interaction, making them slow in dense environments. Biometric technology such as fingerprint and iris recognition prevents proxy fraud but is hygienic and requires cooperation from the user [ 6 ]. Initial face-based attendance systems used handcrafted feature methods. Eigenfaces (PCA decomposition), Fisherfaces (LDA decomposition), and Local Binary Pattern Histograms (LBPH) were the core features in early face recognition systems. These techniques worked well for controlled frontal-face images, but showed lower accuracy in real-life situations such as lighting, pose, and expression. Support Vector Machines (SVM) and k-Nearest Neighbor (k- NN) classifiers were commonly used with these feature extractors, but could not scaled to large institutions [ 7 ]. B. Deep Learnig Architecture for Facial Recognition Deep Convolutional Neural Networks originally emerged in. facial recognition saw the introduction of new advances. DeepFace It was shown in (2014) and the DeepID series that deep learning large numbers of data in hierarchies of features might work. superior to human users in face verification. Models Trained on millions of images, VGGFace and VGGFace2. labeled identity, offer discriminative deep. embeddings that have been adopted to transfer learning in. attendance [ 8 ]. FaceNet (Schroff, et al.) proposed a training strategy based on a triplet loss maximizing the Euclidean distance between embeddings of faces, so that embeddings of the same identity cluster closely and embeddings of different identities are separated. This approach was found to be highly generalizable and used in attendance pipelines. Recent work on angular margin loss functions, such as ArcFace (Additive Angular Margin Loss) and CosFace (Large Margin Cosine Loss) have further reduced the compactness within a class and the separability across classes in the embedding space, producing results on benchmarks such as LFW, IJB-C, and MegaFace [ 9 ]. Vision Transformers (ViT) and their variants are emerging as powerful alternatives to CNNs for face recognition, which use self-attention to model long-range spatial relationships within the facial areas. Hybrid architectures that combine convolutional feature extraction and transformer-based attention have shown superior performance under pose variation and partial occlusion, and thus can be used to observe attendance without constraints [ 10 ]. C. Face Detection Frameworks in Attendance Pipelines Face detection is the first step in attendance recognition pipelines. Multi-task Cascaded Convolutional Networks (MTCNN) is a three-step cascaded technique for face detection and landmark localization. MTCNN can align detected faces perfectly before features are extracted [ 11 ]. The YOLO detectors YOLOv5 and YOLOv8 are used for real-time attendance applications that require simultaneous detection of multiple faces in video. YOLOv8, with its anchor- free detection head and better feature pyramid network, has a higher mean Average Precision (mAP) and lower inference latency than previous YOLO models and is a good choice for smart classrooms and conference rooms. RetinaFace has been proven capable of multitasking detection of facial landmarks and has shown high performance in detecting faces in situations with severe occlusion and scale variation [ 12 ]. D. Transfer Learning and Lightweight Models for Edge Deployment A practical challenge for the admission system is that labeled facial images are scarce and typically for specific user groups and are difficult to track with any advanced edge devices, such as a Raspberry Pi or embedded camera. Transfer learning can be utilized to address this data scarcity problem by using pre-trained models on large publicly available databases (VGGFace2, MS-Celeb-1M) and fine-tuned models on smaller institutional datasets. For example, fine- tuned ArcFace and FaceNet models can recognize individuals with as few as five enrollment images [ 13 ]. In particular, lightweight architectures such as MobileNetV2- based FaceNet models and ShuffleNet-based face recognition models have been optimized for deployment at the edge based on depthwise separable convolutions, knowledge distillation, and quantization. The models have 2– 3% accuracy relative to full-size models, and an inference time under 30 milliseconds on ARM-based edge processors, which enable real-time attendance marking without cloud connectivity [ 14 ]. III. METHODOLOGY A. Face Detection and Preprocessing The attendance recognition pipeline began with video frames continuously captured by the camera positioned in the monitoring area. Each frame passes through a face detection module such as MTCNN or YOLOv8 to locate all faces present in the frame. Face bounding boxes identified are cropped and passed through an alignment stage that uses five facial landmarks (eye centers, nose tip, and mouth corners) to perform affine transformations and normalize each face crop to a standard orientation and scale (typically 112112 or 160160 pixels). Before feature extraction, a histogram equalization or adaptive contrast normalization is applied to compensate for variations in illumination before feature extraction [ 15 ]. B. Feature Extraction Using Deep Learning The pre-processed face crops are fed to a deep feature extraction backbone (ArcFace with ResNet-50/100, FaceNet with InceptionResNetV1, or a lightweight MobileNetV2 version for edge deployment). The backbone outputs a compact, high-dimensional embedding vector (128-D for FaceNet, 512-D for ArcFace) that encodes the discriminative identity information of the face. The embeddings are L2- normalized before being stored or compared, so that similarity can be measured using either the cosine distance or Euclidean distances [ 16 ]. Angular margin-based loss functions (ArcFace, CosFace) used during training keep the embedding space tight and wide cross-class spacing in the embedding space, which improves recognition performance [ 16 ]. C. Matching and Attendance Marking For attendance marking, the extracted embedding of a detected face is compared to a registered database of labeled embeddings by cosine similarity or Euclidean distance. A decision threshold based on balancing the False Acceptance Rate (FAR) and False Rejection Rate (FRR) determines whether a face is a registered identity or an unknown face. If a face is detected above the threshold, the system logs the identity, time, and confidence score to an attendance database. Anti-spoofing modules using depth estimation, texture analysis, or binary classification networks such as FAS-Net are implemented upstream to reject printed photographs, screen replays, and 3D masks before the recognition stage [ 17 ]. D. Multi-Face Simultaneous Recognition In large classes or meetings rooms, multiple faces detected in the same frame need to be processed simultaneously. YOLOv8’s parallel detection feature allows all of the face boxes in a frame to be extracted in one pass; each face is individually processed in the feature extraction and matching step. By selecting multiple embeddings from the database to be processed via matrix (vectorized cosine similarity), the system can recognize multiple faces [ 18 ]. E. Anti-Spoofing and Liveness Detection Spoofing, where someone presents a forged photograph, video replay, or 3D mask to fool the recognition system, is a serious security vulnerability in attendance applications. Anti- spoofing modules employ binary classification networks trained to identify real faces from spoofs using texture data (from LBP), depth maps (from structured light or stereo cameras), and remote photoplethysmography (rPPG). Face Anti-Spoofing (FAS) models with deep learning have been reported as over 98% accurate on benchmarks, and their inclusion as a gate during preprocessing in attendance pipeline has been shown to reduce false acceptance of spoofing attacks [ 19 ]. IV. DATASET AND EVALUATION A. Benchmark Face Recognition Datasets The training and evaluation of deep learning-based facial recognition models for attendance applications require large- scale, high-quality data sources. Important benchmark data sources are: LFW (Labeled Faces in the Wild): 13,233 face images of 5,749 subjects, a set of uncontrolled random images taken from the web, for face verification. VGGFace2: Over 3.3 million images of 9,131 faces with large variation in pose, age, lighting, ethnicity. Widely used as a pre-training component for transfer learning in attendance systems [ 20 ]. MS-Celeb-1M: 10 million images of 100,000 celebrities. Useful for training deep recognition models. IJB-C (IARPA Janus Benchmark-C): Images and videos that show pose, illumination, and occlusion variation. Used to test the recognition robustness in the real world. CASIA-WebFace: 494,414 images of 10,575 subjects. Widely used open-source training data set for face recognition research. B. Attendance-Specific and Institutional Datasets While large public datasets are used for pre-training, fine- tuning and evaluating attendance systems often requires institution-specific datasets. Several studies have created custom datasets to capture student or employee faces in the classroom or office in different poses, distance, and lighting. These datasets typically comprise 5–50 enrollment images per subject and are used to assess recognition performance in the desired environment [ 21 ]. C. Evaluation Metrics Facial recognition attendance systems must be evaluated on five points: Accuracy: Total number of correctly identified individuals in the test set. True Acceptance Rate (TAR) / False Acceptance Rate (FAR): TAR measures correctly identified participants; FAR measures incorrect acceptance of impostors. The TAR@FAR = 0.1% is the standard operating point. False Rejection Rate (FRR): Percent of subjects that were incorrectly rejected by the system. Equal Error Rate (EER): The value of FAR that equals the FRR. A lower EER indicates a better system performance. Proportion of probe faces whose correct matches appear as the top-ranked candidate in the gallery. Processing Latency: Time from frame capture to attendance log entry, critical for deployment in real time. V. RESULT ANALYSIS A. Face Detection Performance The first step in the attendance pipeline is face detection. The most important pieces of literature findings are: MTCNN proves very accurate with frontal and near- frontal face images however recall is poor with severely hidden or profile face images. YOLOv5 can be used to report a mAP of approximately 92–94 on standard face detection benchmarks and inference latency of less than 20ms on general-purpose GPUs. YOLOv8 lies better than YOLOv5 in terms of a rise in mAP 24, and the number of false positives is reduced in large-crowded scenes, thus, it is a better detector to use in multi-face attendance [ 22 ]. TABLE I. Face Detection Performance Comparison Detection Model Dataset mAP (%) Latency (ms) MTCNN WIDER FACE 89.4 45 YOLOv5 WIDER 93.2 18 FACE YOLOv8 WIDER 95.7 12 FACE B. Face Recognition Performance Face recognition models based on deep learning have demonstrated. important enhancements of the traditional methods. Eigenfaces and LBPH have been able to achieve 60–75% accuracy. regulated datasets but worsen significantly with pose and lighting variation. CNN models ( VGGFace2 fine-tuned ) have 88–92% on unrestricted attendance data. ArcFace (ResNet-100 backbone) performs 99.83% on LFW and Rank-1 rate of 96.1% on IJB-C, representing the current state of the art for ensuring identity verification in attendance [ 23 ]. MobileNetV2-FaceNet’s lightweight architecture demonstrates 93–95% accuracy on institutional datasets with a latency of less than 30ms for inference on edge devices allowing real-time deployment without GPUs [ 24 ]. TABLE II. FACE RECOGNITION PERFORMANCE COMPARISON Model Dataset Accuracy (%) Notes LBPH + SVM Institutional 72.4 Classical baseline VGGFace2 (fine-tuned) Institutional 91.3 Transfer learning FaceNet LFW 99.65 Triplet loss ArcFace (ResNet-50) LFW 99.77 Angular margin loss ArcFace (ResNet-100) IJB-C 96.1 (Rank-1) SOTA performance MobileNetV2- FaceNet Institutional 94.2 Edge deployment C. End-to-End Attendance System Performance When evaluating the automated attendance pipeline from face detection, recognition, anti-spoofing, and database integration, several studies report: Systems using MTCNN + FaceNet recognize the subjects end-to-end with accuracy of about 90–93% in controlled classroom environments. Use of YOLOv8 + ArcFace systems achieves 95%–97% attendance marking accuracy in real-time multi-face scenarios. Attentional fusion of multiple enrollees (rather than enrollment in a single image) improves recognition accuracy by 2–4% for persons of varied appearance [ 25 ]. TABLE III. END-TO-END ATTENDANCE SYSTEM PERFORMANCE System Pipeline Environment Accuracy (%) MTCNN + FaceNet Controlled classroom 92.1 MTCNN + ArcFace Office environment 94.8 YOLOv8 + ArcFace Multi-face classroom 96.3 YOLOv8 + ArcFace + Anti- Spoof Real-world deployment 97.1 D. Confusion Analysis and Error Insights Analyses of the pattern of misclassification by system indicate that across the systems reviewed, the patterns are similar: Among the kinds of errors, 40–60% are false rejections of subjects whose appearance has been significantly altered because of hairstyle, glasses, or mask. False acceptance errors are rare in well-tuned systems (FAR 0.5%) but increase substantially when anti-spoofing modules are missing. Demographic differences in error rates (females and darker skinned subjects have a higher FRR) are consistently observed in systems that are trained on mostly non-diverse datasets, such as early versions of MS-Celeb-1M [ 26 ]. Multi-face scenarios with a significant amount of room between subjects lead to a combination of the detection failure and recognition error, and YOLOv8 has a lower rate of these failures than MTCNN. TABLE IV. SUMMARY ACCURACY COMPARISON OF ATTENDANCE RECOGNITION APPROACHES Method Dataset/Environment Accuracy (%) Notes LBPH Controlled institutional 72.4 Classical baseline VGGFace2 + SVM Institutional DB 91.3 Transfer learning FaceNet (MTCNN) LFW 99.65 Triplet loss ArcFace (ResNet- 50) Classroom real-time 95.4 Angular margin YOLOv8 + ArcFace Multi-face classroom 96.3 Real- time pipeline YOLOv8 + ArcFace + FAS Real-world deployment 97.1 With anti- spoofing VI. APPLICATIONS Facial recognition attendance systems have been used in a range of different institutional and organizational contexts. Educational Institutions: Automated student attendance marking in the classroom and examination hall removes the need for proxy attendance, reduces administration time, and provides real-time attendance reports to faculty [ 27 ]. When students are absent without an alert, the LMS could automatically set up alerts that reflect this and generate compliance reports for accreditation [ 28 ]. Corporate and Enterprise: Employee attendance and access controls based on facial recognition can reduce manual entry, interface with payroll to handle time-and-attendance tasks, and audit employee compliance with labor regulations. Hybrid work environments can benefit from virtual attendance verification using cameras in video conferences. Hospitals: Facial recognition allows hospitals to track staff attendance in real time and prevents unauthorized entry to hospital areas. Facial recognition patient identification systems can help prevent medical record mismatches and increase patient safety in healthcare. Government and Military: Facial recognition is utilized for access and attendance verification in high-security areas such as military bases, which are protected from spoofing. For large-scale events or surveillance of attendance at public venues, deep learning is used to analyze crowds and provide security [ 28 ]. Examination and Certification Centers: Anti-impersonation systems in exam rooms use continuous facial recognition to make sure that the registered candidate is actually there during the exam. These systems replace the manual spot checks of invigilators by surveillance robots. VII. ETHICAL CONSIDERATIONS Face recognition attendance systems are gaining in popularity in schools, and it poses a number of ethical, legal, and social problems: Privacy and Biometric Data Protection: Facial images and derived embeddings are sensitive biometric data under GDPR, India's Digital Personal Data Protection Act (DPDPA) 2023, and the EU AI Act (2024). Institutions using facial recognition attendance systems should have the consent of all users, limit the use of information to ensure privacy, limit the use of data to minimal amounts, and store biometric templates in a secure encrypted format. Real-time facial recognition in public spaces is considered a high-risk application of artificial intelligence and must comply with the EU AI Act [ 29 ]. Algorithmic Fairness to Demographic Equity: The research shows that the algorithm is biased. trained models on facial recognition trained on datasets that are demographically imbalanced reject erroneously. A higher number of women and ethnic minorities compared to men. For attendance this might bring about systems that are trained in various institutions. type of inaccurate rejection on some groups of demographics. of technology discrimination. Fairness-aware training, fine-tuning of data sets and performance. All should be evaluated on demographic bases. mitigations [ 30 ]. Anti-Spoofing and Security: The option to be affected by presenters via printed images, video recording playbacks, 3D masks, and increasingly more with deep fake synthetic faces introduces a security vulnerability to attendance systems. The attendance system must have strong liveness detection modules as a standard feature of the system to safeguard against attack on attendance records. Transparency and Explainability: The program participants should be aware of how their biometric data are collected, stored, and utilized in the automated decision-making process. To generate institutional trust and regulatory adherence, explainable AI techniques are required, privacy warnings, and readily available opt-out or objection. Psychological and Surveillance Issues: The existence of a classroom with face recognition through constant monitoring increases. concerns over monitoring students, whether possible or not. freedom of speech is curtailed, and regular assembling of. pupil biometrics data in schools. Educational institutions need to strike a balance with the efficiency benefits of. increased biometric data gathering at the risk of damage. the intellectual and civil freedoms of students. VIII. FUTURE RESEARCH DIRECTIONS New underexplored studies on deep learning-based facial recognition. attendance tracking must support the following issues: Strong Identification at Hard Occlusion: Training. can be used to detect faces partially in their architectures. obstructed by surgical masks, scarves, sunglasses or other. covers is critical towards attendance after a pandemic. The possible lines of research in occlusion-robust models. with partial face reconstruction or regions of attention. masking are being investigated. Lightweight of Edge and Energy-Efficient Architectures. Implementation: Modeling models to high level of accuracy. that can fit the computation and power constraints of edge Status: devices (NVIDIA Jetson Nano, Raspberry Pi 5, or others) The opening of ARM Cortex-M series microcontrollers) will become accessible. establishment of attendance systems to the needy. Federated Learning on Privacy-Protected Training: Facial recognition Learners Federated learning enables two or more institutions to learn facial recognition models in combination without exchanging information on face, thereby satisfying privacy needs and enjoying a host of training strategies. Individual data in training might be also safeguarded by separate privacy measures. Cross-Domain and Cross-Environment generalization: When models are trained in a specific environment and deployed to different settings with different camera parameters, different lighting conditions, or demographics, they can lose accuracy. Domains adaptation and synthetic data augmentation can be used to achieve cross-domain generalization with the aid of generative models. Ongoing Education on Adapting Databases: Attendance databases are dynamic, as new users are added, look different as people grow old and some frequenters drop out. One of the most crucial elements is to make sure that a model can be continuously learned without eliminating identities that used to be in the model. Deepfake Detection: With more believable synthetic faces generated by generative AI, attendance systems should be able to recognize a deepfake just like typical anti-spoofing ability to ensure a foiled presentation in the future. Explainable and Auditable Recognition Systems: Designing a visualization of interpretable attention features that indicate which regions of the face were used to predict whether or not the facial expression of a person was recognized will enhance clarity of the systems, help bias auditing and make regulatory compliance reviews easier. IX. CONCLUSION Facial recognition based on deep learning has radically transformed the way attendance is monitored by providing non-intrusive, real-time, and accurate identification for students in all types of institutions. This review examines the evolution from handcrafted feature learning to deep metric learning using CNNs, ResNets, FaceNet, ArcFace, CosFace, Vision Transformers, and many others. For face detection, the first step of any attendance pipeline, MTCNN and YOLO-based algorithms, particularly YOLOv8, can simultaneously detect multiple faces at high frame rates and mAP. In recognition, angular margin models such as ArcFace set the standard for identity verification accuracy, while lighter MobileNetV2-based models make edge deployment feasible without sacrificing accuracy. End to-end attendance pipelines with YOLOv8 detection, ArcFace recognition, and integrated anti-spoofing capabilities typically achieve higher than 96–97% accuracy in real world applications. These have been made, but there are still many challenges that should be made addressed. Stability in the cases of a low- vision, in cases of occlusion, low- light visibility, etc. extreme is worsened under open-air conditions. The demographic bias may be the cause of power in the pre-training data sets and the aspect of unfairness, that should be fixed. Biometric data should be administrated in the privacy and law, such as GDPR and EU AI Act. Lightweight deployment at the edge, federated training and privacy enforcement training, wide-range cross-environment generalization, continuous learning of dynamically changing databases, and explainable AI to understand errors with a clear vision are the areas that will allow making more progress in this direction in the future. Not only will high- tech expertise and ethical deployment and compliance be necessary to provide reliable, scalable, and fair automated attendance monitoring, but also ethical usage and enforcement. Declarations Funding The authors declare that no funding was received for this research. Author Contribution P.K. and R.R. conceptualized the study and designed the overall research framework. R.G. and M.K. conducted the literature review and analyzed existing deep learning models and facial recognition techniques. S. contributed to the methodology design, including face detection, feature extraction, and system pipeline development. A.K. assisted in data interpretation, comparative analysis, and preparation of tables and performance evaluation.P.K. and R.R. drafted the main manuscript text. R.G. and M.K. contributed to writing specific sections, including literature review and results analysis. S. reviewed the technical accuracy and refined the methodology section. A.K. prepared figures, formatting, and assisted in final editing.All authors reviewed, revised, and approved the final version of the manuscript. References Thalor MA, Gaikwad OS (2023) Facial Recognition Attendance Monitoring System using Deep Learning Techniques, Int. J. Integrated Sci. Technol., vol. 1, no. 6, pp. 841–848, Dec Fadel NEL (2025) Facial Recognition Algorithms: A Systematic Literature Review, J. Imaging, vol. 11, no. 2, p. 58, Feb Saranya S, Chandru M, Ragul S, Kafeel SA (2024) Facial Recognition Attendance Monitoring System Using Deep Learning with YOLOv8. Int J Adv Res Sci Eng Manag (IJARESM) Aidana A et al (2025) Robust Face Recognition Under Challenging Conditions: A Comprehensive Review of Deep Learning Methods, Appl. Sci., vol. 15, no. 17, p. 9390, Aug Wang Y, Wu L, Zhou X, Fu H (2024) Beyond Surveillance: Privacy, Ethics, and Regulations in Face Recognition Technology, Front. Big Data, vol. 7, p. 1337465, Jul Painuly K, Bisht Y, Vaidya H, Kapruwan A, Gupta R (2024) Efficient Real-Time Face Recognition-Based Attendance System with Deep Learning Algorithms, in Proc. 2024 4th Int. Conf. Adv. Comput. Commun. Inform. Technol. (ACCIT), IEEE Verma A, Gupta R, Sharma S (2023) Automated Attendance Monitoring System Using Facial Recognition, in Proc. 2023 Int. Conf. Emerging Trends Comput. Sci. Inf. Technol Haq MU, Sethi MAJ, Ahmad S et al (2025) A Comprehensive Review of Face Detection/Recognition Algorithms and Competitive Datasets. Comput Mater Continua 84(1):1–24 Juwanda RC, Alunjati FA, Elviani U, Hidayat FP (2024) Comparative Analysis of FaceNet and ArcFace in Minimizing False Positives for Enhanced Access Control Security, in Proc. 2024 Int. Conf. ICT Smart Soc. (ICISS), IEEE, Sep. pp. 1–6 Mustafa S, Wang J, Liu H (Jun. 2024) Towards Efficient and Robust Face Recognition Through Attention-Integrated Multi-Level CNN, Multimedia Tools Appl. Springer Nature Karamizadeh S, Chaeikar SS, Salarian H (2025) Combining MTCNN and Enhanced FaceNet with Adaptive Feature Fusion for Robust Face Recognition, Technologies, Oct Chansaeng T, Jarupunphol P (2024) Attendance System Optimization through Deep Learning Face Recognition, Int. J. Comput. Digit. Syst., vol. 15, no. 1, pp. 1527–1540, Apr Dang H (2023) Int Efficient Deep Learning Approach for Facial Recognition Using Improved FaceNet Based on MobileNetV2, in Proc. Conf. Comput. Sci. Inf. Technol., 2023 Jarupunphol P, Chansaeng T (2024) An Automated Face Detection and Recognition for Class. Attendance Using IoT and Edge Computing, Fernando R, Athauda H (2024) Image Processing Based Real-Time Online Attendance Monitoring System Using Facial Recognition, in Proc. 2024 Int. Conf. Image Process. Robot. (ICIPRoB), IEEE Chin KC et al (2024) Deep Learning Based Attendance Check System, in Proc. 2024 9th Int. Conf. Intell. Inf. Technol. (ICIIT), ACM Balakannan SP, Sravani K, Jeyapriya S, Fathima, Rosini AIB (2024) Attendance System Using Machine Learning, in Proc. 2024 Int. Conf. Electr. Electron. Comput. Technol. (ICEECT), IEEE Talukder OM, Dakhil A, Sheikh M (2024) Automation of Surveillance Systems Using Deep Learning and Facial Recognition, J. Inf. Syst. Eng. Manag., vol. 10, no. 35s Ponnusamy M et al (2024) Enhancing Security in Public Spaces Through Generative Adversarial Networks (GANs), J. Inf. Syst. Eng. Manag., vol. 10, no. 35s Sadak MS (2025) Face Recognition Technologies: A Comprehensive Review. Int J Sci Basic Appl Res (IJSBAR) 77(1):333–345 Abraham CS, Dev D, Patnaik R (2023) Int Challenges in Facial Feature- Based Recognition: Glasses, Beards, and Demographic Variability, in Proc. Conf. Comput. Commun. Inform., 2023 Bose S, Kumar M, Patel R (2024) Efficient Deep Models for Real-Time Face Recognition. IEEE Trans Inf Forensics Secur Anonymous (2025) Evaluation of Deep Learning Methods in Face Recognition. Int J Sci Adv Technol (IJSAT) Raj B, Kumar S, Mehra P (2023) Comparative Analysis of Eigenfaces, Fisherfaces, and LBPH Under Variable Illumination for Attendance Systems. Int J Comput Appl 185(42):1–7 Anonymous AI-P (2025) Facial Recognition Attendance System, Int. J. Res. Sci. Innovation (IJRSI), vol. 12, no. 9, pp. 3902–3912, Sep Fadel N et al (2025) Facial Recognition Algorithms: Systematic Literature Review Focusing on Occlusion and Bias. J Imaging, 11, 2 Anonymous (2025) 50 Years of Automated Face Recognition, arXiv preprint arXiv:2505.24247 Anonymous (Jun. 2025) Global Perspectives on Regulating Facial Recognition Technology Utilization for Criminal Justice, Global Public Policy Governance. Springer Nature Aidana A et al (2025) Deep Learning and Facial Biometrics: Compliance with EU AI Act Requirements. Appl Sci Verma A, Kumar S (2024) Fairness and Bias Evaluation in Institutional Facial Recognition Attendance Systems, in Proc. 2024 IEEE Int. Conf. Biometrics Theory Appl. Syst. (BTAS), IEEE Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9504761","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":629918660,"identity":"829cad0b-4e9b-44e0-ace1-4b3085b5cbea","order_by":0,"name":"Raj Ranjan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+0lEQVRIiWNgGAWjYLCCBwwMzGzM7AcOfABy2NgJqmdmYEgAknzsPIkPZ4C0MBOphUGOn8HYmAcqgBeYs58/+CBxjx070PA0aZtf2+T5mBkYP3zMwa3FsieZ2SDhWTLQL4zHpHP7bhu2MTMwS87chluLwYFkNomEA8zMYFtye24zArWwMfPi03L+MfuPhAP1IC1m0pY9t+0Ja7mRzMaQcOAwSIuxMcOP24kEtVjOeGwMdNhxoBZgIPc23E5uY2ZsxusXc/7Ehx8+HKhOlu8/fuDAjz+3bee3Nx/88BGfw6B0MphkbAOTDbjVI2mxg1B/8CoeBaNgFIyCEQoA0zhKiWp7OxQAAAAASUVORK5CYII=","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":true,"prefix":"","firstName":"Raj","middleName":"","lastName":"Ranjan","suffix":""},{"id":629918661,"identity":"0a1f1a48-25e9-4780-98dc-73c85a12fb98","order_by":1,"name":"Sureshwati MCA","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Sureshwati","middleName":"","lastName":"MCA","suffix":""},{"id":629918662,"identity":"5d65bbcb-fa43-4eb6-9f16-1fc33bc9bed0","order_by":2,"name":"Manish Kumar","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Manish","middleName":"","lastName":"Kumar","suffix":""},{"id":629918663,"identity":"02d6bf2f-3316-41e3-a7b8-fa337d04dc92","order_by":3,"name":"Rahul Gandhi","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Rahul","middleName":"","lastName":"Gandhi","suffix":""},{"id":629918664,"identity":"fc663c35-6eed-4f55-8644-ce46275e4b64","order_by":4,"name":"Pawan Kumar","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Pawan","middleName":"","lastName":"Kumar","suffix":""},{"id":629918665,"identity":"a89e5012-b1de-43d7-aca8-b20e1495820e","order_by":5,"name":"Harendra Singh","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Harendra","middleName":"","lastName":"Singh","suffix":""},{"id":629918666,"identity":"3e769ee5-c93a-443e-91eb-489f39cf34c4","order_by":6,"name":"Aditya Kumar","email":"","orcid":"","institution":"Greater Noida Institute of Technology(Engg. Institute)","correspondingAuthor":false,"prefix":"","firstName":"Aditya","middleName":"","lastName":"Kumar","suffix":""}],"badges":[],"createdAt":"2026-04-23 09:23:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9504761/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9504761/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108075574,"identity":"cb9a0143-fb7a-4e7d-908d-2be4a298044d","added_by":"auto","created_at":"2026-04-29 06:46:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":93578,"visible":true,"origin":"","legend":"\u003cp\u003eFace Detaction and Preprocessing pipeline\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9504761/v1/90bf9b9dc1b4b8a7e3f0d9a4.png"},{"id":108181465,"identity":"da880827-40a1-4507-a742-cfd472237e69","added_by":"auto","created_at":"2026-04-30 08:58:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":98163,"visible":true,"origin":"","legend":"\u003cp\u003eDeep Feature Extraction Workflow\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9504761/v1/3bbe12f4250240144255fc6a.png"},{"id":108181810,"identity":"b26a93a3-748c-4223-852d-ef6f51b68f0a","added_by":"auto","created_at":"2026-04-30 08:58:56","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":7443,"visible":true,"origin":"","legend":"\u003cp\u003eMulti-Face Simultaneous Recognition and Attendance Marking Pipeline\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9504761/v1/aad4641db4e25ad549e4e8f8.jpg"},{"id":108491923,"identity":"e4d61b48-5e4c-47ca-9375-0046f60a9a3d","added_by":"auto","created_at":"2026-05-05 09:56:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":453844,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9504761/v1/6c9e3716-12d1-49c2-b092-7b4f22a51745.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Comprehensive Review of Deep Learning- Driven Facial Recognition Frameworks for Intelligent and Fully Automated Attendance Monitoring Systems","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eThe advent of artificial intelligence and computer. vision technologies has revolutionized the ways. businesses check their identity and access control. One of the attendance is by far the largest example of this new era. monitoring. The facial recognition technology (FRT) enables. organizations to determine people in real-time without the requirement.\u003c/p\u003e\n\u003cp\u003eproximity cards, or manually entering names into a roll call box [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e]. FRT is much more safe than finger or iris scanning. Face recognition can be used in large-scale environments such as universities, office buildings, and health canter’s.\u003c/p\u003e\n\u003cp\u003eDuring the past decade, deep learning has revolutionized facial recognition. CNNs, ResNets, FaceNet, ArcFace, and Vision Transformers consistently achieve near-human or superior accuracy on standard benchmark datasets [\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e]. From handcrafted features like Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) to deep learned embeddings, we have been able to generalize across lighting, pose, and occlusion. Meanwhile, Multi-task Cascaded Convolutional Networks (MTCNN) and YOLO detectors have enabled simultaneous multi-face detection in real time video, which is necessary to accurately track attendance in crowded classrooms or meetings [\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eDespite these advances, deep learning-based attendance systems can still be difficult to deploy on an institutional scale. For instance, facial recognition with partial face occlusion due to masks or glasses, varying lighting, morphology, and low resolution input data can reduce performance dramatically [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e]. Privacy issues with biometric data storage, protection against spoofing, and requirements for compliance with regulations such as GDPR and EU AI Act (2024) continue to hinder large- scale application [\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e]. Review of the literature reveals many independent applications of deep learning, but few recent and current review studies that are looking at the architectures, face detection pipelines, benchmark datasets, real world deployment constraints, ethical considerations, and performance trade-offs in attendance monitoring. This paper attempts to fill this gap by providing a systematic review of deep learning-based facial recognition applied to attendance monitoring. It covers the entire process of recognition from face detection and preprocessing to feature extraction, matching, and system integration.\u003c/p\u003e\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n"},{"header":"II. LITERATURE REVIEW AND RELATED WORK","content":"\u003cp\u003e\u003cem\u003eA. Traditional and Classical Approaches to Attendance Monitoring\u003c/em\u003e\u003c/p\u003e\u003cp\u003eInitial attendance systems were based on proximity technologies such as RFID cards, barcode scanning, and PIN- based entry. These are functional, but allow for proxy attendance (when someone marks one’s attendance for another) and require physical interaction, making them slow in dense environments. Biometric technology such as fingerprint and iris recognition prevents proxy fraud but is hygienic and requires cooperation from the user [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eInitial face-based attendance systems used handcrafted feature methods. Eigenfaces (PCA decomposition), Fisherfaces (LDA decomposition), and Local Binary Pattern Histograms (LBPH) were the core features in early face recognition systems. These techniques worked well for controlled frontal-face images, but showed lower accuracy in real-life situations such as lighting, pose, and expression. Support Vector Machines (SVM) and k-Nearest Neighbor (k- NN) classifiers were commonly used with these feature extractors, but could not scaled to large institutions [\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eB. Deep Learnig Architecture for Facial Recognition\u003c/em\u003e\u003c/p\u003e\u003cp\u003eDeep Convolutional Neural Networks originally emerged in. facial recognition saw the introduction of new advances. DeepFace It was shown in (2014) and the DeepID series that deep learning large numbers of data in hierarchies of features might work. superior to human users in face verification. Models Trained on millions of images, VGGFace and VGGFace2. labeled identity, offer discriminative deep. embeddings that have been adopted to transfer learning in. attendance [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFaceNet (Schroff, et al.) proposed a training strategy based on a triplet loss maximizing the Euclidean distance between embeddings of faces, so that embeddings of the same identity cluster closely and embeddings of different identities are separated. This approach was found to be highly generalizable and used in attendance pipelines. Recent work on angular margin loss functions, such as ArcFace (Additive Angular Margin Loss) and CosFace (Large Margin Cosine Loss) have further reduced the compactness within a class and the separability across classes in the embedding space, producing results on benchmarks such as LFW, IJB-C, and MegaFace [\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eVision Transformers (ViT) and their variants are emerging as powerful alternatives to CNNs for face recognition, which use self-attention to model long-range spatial relationships within the facial areas. Hybrid architectures that combine convolutional feature extraction and transformer-based attention have shown superior performance under pose variation and partial occlusion, and thus can be used to observe attendance without constraints [\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eC. Face Detection Frameworks in Attendance Pipelines\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFace detection is the first step in attendance recognition pipelines. Multi-task Cascaded Convolutional Networks (MTCNN) is a three-step cascaded technique\u003c/p\u003e\u003cp\u003efor face detection and landmark localization. MTCNN can align detected faces perfectly before features are extracted [\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe YOLO detectors YOLOv5 and YOLOv8 are used for real-time attendance applications that require simultaneous detection of multiple faces in video. YOLOv8, with its anchor- free detection head and better feature pyramid network, has a higher mean Average Precision (mAP) and lower inference latency than previous YOLO models and is a good choice for smart classrooms and conference rooms. RetinaFace has been proven capable of multitasking detection of facial landmarks and has shown high performance in detecting faces in situations with severe occlusion and scale variation [\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eD. Transfer Learning and Lightweight Models for Edge Deployment\u003c/em\u003e\u003c/p\u003e\u003cp\u003eA practical challenge for the admission system is that labeled facial images are scarce and typically for specific user groups and are difficult to track with any advanced edge devices, such as a Raspberry Pi or embedded camera. Transfer learning can be utilized to address this data scarcity problem by using pre-trained models on large publicly available databases (VGGFace2, MS-Celeb-1M) and fine-tuned models on smaller institutional datasets. For example, fine- tuned ArcFace and FaceNet models can recognize individuals with as few as five enrollment images [\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn particular, lightweight architectures such as MobileNetV2- based FaceNet models and ShuffleNet-based face recognition models have been optimized for deployment at the edge based on depthwise separable convolutions, knowledge distillation, and quantization. The models have 2– 3% accuracy relative to full-size models, and an inference time under 30 milliseconds on ARM-based edge processors, which enable real-time attendance marking without cloud connectivity [\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e"},{"header":"III. METHODOLOGY","content":"\u003cp\u003e\u003cem\u003eA. Face Detection and Preprocessing\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe attendance recognition pipeline began with video frames continuously captured by the camera positioned in the monitoring area. Each frame passes through a face detection module such as MTCNN or YOLOv8 to locate all faces present in the frame. Face bounding boxes identified are cropped and passed through an alignment stage that uses five facial landmarks (eye centers, nose tip, and mouth corners) to perform affine transformations and normalize each face crop to a standard orientation and scale (typically 112112 or 160160 pixels). Before feature extraction, a histogram equalization or adaptive contrast normalization is applied to compensate for variations in illumination before feature extraction [\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eB.\u0026nbsp;Feature Extraction Using Deep Learning\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe pre-processed face crops are fed to a deep feature extraction backbone (ArcFace with ResNet-50/100, FaceNet with InceptionResNetV1, or a lightweight MobileNetV2 version for edge deployment). The backbone outputs a compact, high-dimensional embedding vector (128-D for FaceNet, 512-D for ArcFace) that encodes the discriminative identity information of the face. The embeddings are L2- normalized before being stored or compared, so that similarity can be measured using either the cosine distance or Euclidean distances [\u003cspan class=\"CitationRef\"\u003e16\u003c/span\u003e]. Angular margin-based loss functions (ArcFace, CosFace) used during training keep the embedding space tight and wide cross-class spacing in the embedding space, which improves recognition performance [\u003cspan class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eC. Matching and Attendance Marking\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFor attendance marking, the extracted embedding of a detected face is compared to a registered database of labeled embeddings by cosine similarity or Euclidean distance. A decision threshold based on balancing the False Acceptance Rate (FAR) and False Rejection Rate (FRR) determines whether a face is a registered identity or an unknown face. If a face is detected above the threshold, the system logs the identity, time, and confidence score to an attendance database. Anti-spoofing modules using depth estimation, texture analysis, or binary classification networks such as FAS-Net are implemented upstream to reject printed photographs, screen replays, and 3D masks before the recognition stage [\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eD. Multi-Face Simultaneous Recognition\u003c/em\u003e\u003c/p\u003e\u003cp\u003eIn large classes or meetings rooms, multiple faces detected in the same frame need to be processed simultaneously. YOLOv8’s parallel detection feature allows all of the face boxes in a frame to be extracted in one pass; each face is individually processed in the feature extraction and matching step. By selecting multiple embeddings from the database to be processed via matrix (vectorized cosine similarity), the system can recognize multiple faces [\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eE.\u0026nbsp;Anti-Spoofing and Liveness Detection\u003c/em\u003e\u003c/p\u003e\u003cp\u003eSpoofing, where someone presents a forged photograph, video replay, or 3D mask to fool the recognition system, is a serious security vulnerability in attendance applications. Anti- spoofing modules employ binary classification networks trained to identify real faces from spoofs using texture data (from LBP), depth maps (from structured light or stereo cameras), and remote photoplethysmography (rPPG). Face Anti-Spoofing (FAS) models with deep learning have been reported as over 98% accurate on benchmarks, and their inclusion as a gate during preprocessing in attendance pipeline has been shown to reduce false acceptance of spoofing attacks [\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e"},{"header":"IV. DATASET AND EVALUATION","content":"\u003cp\u003e\u003cem\u003eA. Benchmark Face Recognition Datasets\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe training and evaluation of deep learning-based facial recognition models for attendance applications require large- scale, high-quality data sources. Important benchmark data sources are:\u003c/p\u003e\u003cp\u003eLFW (Labeled Faces in the Wild): 13,233 face images of 5,749 subjects, a set of uncontrolled random images taken from the web, for face verification.\u003c/p\u003e\u003cp\u003eVGGFace2: Over 3.3\u0026nbsp;million images of 9,131 faces with large variation in pose, age, lighting, ethnicity. Widely used as a pre-training component for transfer learning in attendance systems [\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMS-Celeb-1M: 10\u0026nbsp;million images of 100,000 celebrities. Useful for training deep recognition models. IJB-C (IARPA Janus Benchmark-C): Images and videos that show pose, illumination, and occlusion variation. Used to test the recognition robustness in the real world. CASIA-WebFace: 494,414 images of 10,575 subjects. Widely used open-source training data set for face recognition research.\u003c/p\u003e\u003cp\u003e\u003cem\u003eB. Attendance-Specific and Institutional Datasets\u003c/em\u003e\u003c/p\u003e\u003cp\u003eWhile large public datasets are used for pre-training, fine- tuning and evaluating attendance systems often requires institution-specific datasets. Several studies have created custom datasets to capture student or employee faces in the\u003c/p\u003e\u003cp\u003eclassroom or office in different poses, distance, and lighting. These datasets typically comprise 5–50 enrollment images per subject and are used to assess recognition performance in the desired environment [\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cem\u003eC. Evaluation Metrics\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFacial recognition attendance systems must be evaluated on five points:\u003c/p\u003e\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eAccuracy: Total number of correctly identified individuals in the test set.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eTrue Acceptance Rate (TAR) / False Acceptance Rate (FAR): TAR measures correctly identified participants; FAR measures incorrect acceptance of impostors. The TAR@FAR = 0.1% is the standard operating point.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eFalse Rejection Rate (FRR): Percent of subjects that were incorrectly rejected by the system.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEqual Error Rate (EER): The value of FAR that equals the FRR. A lower EER indicates a better system performance. Proportion of probe faces whose correct matches appear as the top-ranked candidate in the gallery.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eProcessing Latency: Time from frame capture to attendance log entry, critical for deployment in real time.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"V. RESULT ANALYSIS","content":"\u003cp\u003e\u003cem\u003eA. Face Detection Performance\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe first step in the attendance pipeline is face detection. The most important pieces of literature findings are:\u003c/p\u003e\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eMTCNN proves very accurate with frontal and near- frontal face images however recall is poor with severely hidden or profile face images.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eYOLOv5 can be used to report a mAP of approximately 92–94 on standard face detection benchmarks and inference latency of less than 20ms on general-purpose GPUs.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eYOLOv8 lies better than YOLOv5 in terms of a rise in mAP 24, and the number of false positives is reduced in large-crowded scenes, thus, it is a better detector to use in multi-face attendance [\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\u003cp\u003e\u003cspan class=\"SmallCaps\"\u003eTABLE I. Face Detection Performance Comparison\u003c/span\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Taba\" border=\"1\"\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eDetection Model\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eDataset\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003emAP (%)\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eLatency (ms)\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMTCNN\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWIDER FACE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e89.4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e45\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWIDER\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e93.2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e18\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFACE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv8\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWIDER\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e95.7\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e12\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFACE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\u003cp\u003e\u003cem\u003eB. Face Recognition Performance\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFace recognition models based on deep learning have demonstrated. important enhancements of the traditional methods. Eigenfaces and LBPH have been able to achieve 60–75% accuracy. regulated datasets but worsen significantly with pose and lighting variation. CNN models ( VGGFace2 fine-tuned ) have 88–92% on unrestricted attendance data. ArcFace (ResNet-100 backbone) performs 99.83% on LFW\u003c/p\u003e\u003cp\u003eand Rank-1 rate of 96.1% on IJB-C, representing the current state of the art for ensuring identity verification in attendance [\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMobileNetV2-FaceNet’s lightweight architecture demonstrates 93–95% accuracy on institutional datasets with a latency of less than 30ms for inference on edge devices allowing real-time deployment without GPUs [\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTABLE II. FACE RECOGNITION PERFORMANCE COMPARISON\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tabb\" border=\"1\"\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eDataset\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy (%)\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eNotes\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLBPH + SVM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInstitutional\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e72.4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eClassical baseline\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eVGGFace2 (fine-tuned)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInstitutional\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e91.3\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTransfer learning\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFaceNet\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLFW\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e99.65\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTriplet loss\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eArcFace (ResNet-50)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLFW\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e99.77\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAngular margin loss\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eArcFace (ResNet-100)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eIJB-C\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e96.1\u003c/p\u003e\n\u003cp\u003e(Rank-1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eSOTA\u003c/p\u003e\n\u003cp\u003eperformance\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMobileNetV2- FaceNet\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInstitutional\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e94.2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEdge deployment\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\u003cp\u003e\u003cem\u003eC. End-to-End Attendance System Performance\u003c/em\u003e\u003c/p\u003e\u003cp\u003eWhen evaluating the automated attendance pipeline from face detection, recognition, anti-spoofing, and database integration, several studies report:\u003c/p\u003e\u003cp\u003eSystems using MTCNN + FaceNet recognize the subjects end-to-end with accuracy of about 90–93% in controlled classroom environments.\u003c/p\u003e\u003cp\u003eUse of YOLOv8 + ArcFace systems achieves 95%–97% attendance marking accuracy in real-time multi-face scenarios.\u003c/p\u003e\u003cp\u003eAttentional fusion of multiple enrollees (rather than enrollment in a single image) improves recognition accuracy by 2–4% for persons of varied appearance [\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTABLE III. END-TO-END ATTENDANCE SYSTEM PERFORMANCE\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tabc\" border=\"1\"\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eSystem Pipeline\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eEnvironment\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy (%)\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMTCNN +\u003c/p\u003e\n\u003cp\u003eFaceNet\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eControlled classroom\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e92.1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMTCNN +\u003c/p\u003e\n\u003cp\u003eArcFace\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOffice environment\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e94.8\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv8 +\u003c/p\u003e\n\u003cp\u003eArcFace\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMulti-face classroom\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e96.3\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv8 +\u003c/p\u003e\n\u003cp\u003eArcFace + Anti- Spoof\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eReal-world deployment\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e97.1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\u003cp\u003e\u003cem\u003eD. Confusion Analysis and Error Insights\u003c/em\u003e\u003c/p\u003e\u003cp\u003eAnalyses of the pattern of misclassification by system indicate that across the systems reviewed, the patterns are similar:\u003c/p\u003e\u003cp\u003eAmong the kinds of errors, 40–60% are false rejections of subjects whose appearance has been significantly altered because of hairstyle, glasses, or mask. False acceptance errors are rare in well-tuned systems (FAR 0.5%) but increase substantially when anti-spoofing modules are missing.\u003c/p\u003e\u003cp\u003eDemographic differences in error rates (females and darker skinned subjects have a higher FRR) are consistently observed in systems that are trained on mostly non-diverse datasets, such as early versions of MS-Celeb-1M [\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMulti-face scenarios with a significant amount of room between subjects lead to a combination of the detection failure and recognition error, and YOLOv8 has a lower rate of these failures than MTCNN.\u003c/p\u003e\u003cp\u003eTABLE IV. SUMMARY ACCURACY COMPARISON OF ATTENDANCE RECOGNITION APPROACHES\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tabd\" border=\"1\"\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eMethod\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eDataset/Environment\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy (%)\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eNotes\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLBPH\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eControlled institutional\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e72.4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eClassical baseline\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eVGGFace2\u003c/p\u003e\n\u003cp\u003e+ SVM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInstitutional DB\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e91.3\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTransfer learning\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFaceNet (MTCNN)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLFW\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e99.65\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTriplet loss\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eArcFace (ResNet- 50)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eClassroom real-time\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e95.4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAngular margin\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv8\u003c/p\u003e\n\u003cp\u003e+ ArcFace\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMulti-face classroom\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e96.3\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eReal- time pipeline\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYOLOv8\u003c/p\u003e\n\u003cp\u003e+ ArcFace\u003c/p\u003e\n\u003cp\u003e+ FAS\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eReal-world deployment\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\"\u003e\n\u003cp\u003e97.1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWith anti- spoofing\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e"},{"header":"VI. APPLICATIONS","content":"\u003cp\u003eFacial recognition attendance systems have been used in a range of different institutional and organizational contexts.\u003c/p\u003e \u003cp\u003eEducational Institutions: Automated student attendance marking in the classroom and examination hall removes the need for proxy attendance, reduces administration time, and provides real-time attendance reports to faculty [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. When students are absent without an alert, the LMS could automatically set up alerts that reflect this and generate compliance reports for accreditation [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCorporate and Enterprise: Employee attendance and access controls based on facial recognition can reduce manual entry, interface with payroll to handle time-and-attendance tasks, and audit employee compliance with labor regulations. Hybrid work environments can benefit from virtual attendance verification using cameras in video conferences.\u003c/p\u003e \u003cp\u003eHospitals: Facial recognition allows hospitals to track staff attendance in real time and prevents unauthorized entry to hospital areas. Facial recognition patient identification systems can help prevent medical record mismatches and increase patient safety in healthcare.\u003c/p\u003e \u003cp\u003eGovernment and Military: Facial recognition is utilized for access and attendance verification in high-security areas such as military bases, which are protected from spoofing. For large-scale events or surveillance of attendance at public venues, deep learning is used to analyze crowds and provide security [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eExamination and Certification Centers: Anti-impersonation systems in exam rooms use continuous facial recognition to make sure that the registered candidate is actually there during the exam. These systems replace the manual spot checks of invigilators by surveillance robots.\u003c/p\u003e"},{"header":"VII. ETHICAL CONSIDERATIONS","content":"\u003cp\u003eFace recognition attendance systems are gaining in popularity in schools, and it poses a number of ethical, legal, and social problems:\u003c/p\u003e \u003cp\u003ePrivacy and Biometric Data Protection: Facial images and derived embeddings are sensitive biometric data under GDPR, India's Digital Personal Data Protection Act (DPDPA) 2023, and the EU AI Act (2024). Institutions using facial recognition attendance systems should have the consent of all users, limit the use of information to ensure privacy, limit the use of data to minimal amounts, and store biometric templates in a secure encrypted format. Real-time facial recognition in public spaces is considered a high-risk application of artificial intelligence and must comply with the EU AI Act [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlgorithmic Fairness to Demographic Equity: The research shows that the algorithm is biased. trained models on facial recognition trained on datasets that are demographically imbalanced reject erroneously. A higher number of women and ethnic minorities compared to men. For attendance this might bring about systems that are trained in various institutions. type of inaccurate rejection on some groups of demographics. of technology discrimination. Fairness-aware training, fine-tuning of data sets and performance. All should be evaluated on demographic bases. mitigations [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAnti-Spoofing and Security: The option to be affected by presenters via printed images, video recording playbacks, 3D masks, and increasingly more with deep fake synthetic faces introduces a security vulnerability to attendance systems. The attendance system must have strong liveness detection modules as a standard feature of the system to safeguard against attack on attendance records.\u003c/p\u003e \u003cp\u003eTransparency and Explainability: The program participants should be aware of how their biometric data are collected, stored, and utilized in the automated decision-making process. To generate institutional trust and regulatory adherence, explainable AI techniques are required, privacy warnings, and readily available opt-out or objection.\u003c/p\u003e \u003cp\u003ePsychological and Surveillance Issues: The existence of a classroom with face recognition through constant monitoring increases. concerns over monitoring students, whether possible or not. freedom of speech is curtailed, and regular assembling of. pupil biometrics data in schools. Educational institutions need to strike a balance with the efficiency benefits of. increased biometric data gathering at the risk of damage. the intellectual and civil freedoms of students.\u003c/p\u003e"},{"header":"VIII. FUTURE RESEARCH DIRECTIONS","content":"\u003cp\u003eNew underexplored studies on deep learning-based facial recognition. attendance tracking must support the following issues:\u003c/p\u003e \u003cp\u003eStrong Identification at Hard Occlusion: Training. can be used to detect faces partially in their architectures. obstructed by surgical masks, scarves, sunglasses or other. covers is critical towards attendance after a pandemic. The possible lines of research in occlusion-robust models. with partial face reconstruction or regions of attention. masking are being investigated.\u003c/p\u003e \u003cp\u003eLightweight of Edge and Energy-Efficient Architectures. Implementation: Modeling models to high level of accuracy. that can fit the computation and power constraints of edge Status: devices (NVIDIA Jetson Nano, Raspberry Pi 5, or others) The opening of ARM Cortex-M series microcontrollers) will become accessible. establishment of attendance systems to the needy.\u003c/p\u003e \u003cp\u003eFederated Learning on Privacy-Protected Training: Facial recognition Learners Federated learning enables two or more institutions to learn facial recognition models in combination without exchanging information on face, thereby satisfying privacy needs and enjoying a host of training strategies. Individual data in training might be also safeguarded by separate privacy measures.\u003c/p\u003e \u003cp\u003eCross-Domain and Cross-Environment generalization: When models are trained in a specific environment and deployed to different settings with different camera parameters, different lighting conditions, or demographics, they can lose accuracy. Domains adaptation and synthetic data augmentation can be used to achieve cross-domain generalization with the aid of generative models.\u003c/p\u003e \u003cp\u003eOngoing Education on Adapting Databases: Attendance databases are dynamic, as new users are added, look different as people grow old and some frequenters drop out. One of the most crucial elements is to make sure that a model can be continuously learned without eliminating identities that used to be in the model. Deepfake Detection: With more believable synthetic faces generated by generative AI, attendance systems should be able to recognize a deepfake just like typical anti-spoofing ability to ensure a foiled presentation in the future.\u003c/p\u003e \u003cp\u003eExplainable and Auditable Recognition Systems: Designing a visualization of interpretable attention features that indicate which regions of the face were used to predict whether or not the facial expression of a person was recognized will enhance clarity of the systems, help bias auditing and make regulatory compliance reviews easier.\u003c/p\u003e"},{"header":"IX. CONCLUSION","content":"\u003cp\u003eFacial recognition based on deep learning has radically transformed the way attendance is monitored by providing non-intrusive, real-time, and accurate identification for students in all types of institutions. This review examines the evolution from handcrafted feature learning to deep metric learning using CNNs, ResNets, FaceNet, ArcFace, CosFace, Vision Transformers, and many others.\u003c/p\u003e \u003cp\u003eFor face detection, the first step of any attendance pipeline, MTCNN and YOLO-based algorithms, particularly YOLOv8, can simultaneously detect multiple faces at high frame rates and mAP. In recognition, angular margin models such as ArcFace set the standard for identity verification accuracy, while lighter MobileNetV2-based models make edge deployment feasible without sacrificing accuracy. End to-end attendance pipelines with YOLOv8 detection, ArcFace recognition, and integrated anti-spoofing capabilities typically achieve higher than 96\u0026ndash;97% accuracy in real world applications.\u003c/p\u003e \u003cp\u003eThese have been made, but there are still many challenges that should be made addressed. Stability in the cases of a low- vision, in cases of occlusion, low- light visibility, etc. extreme is worsened under open-air conditions. The demographic bias may be the cause of power in the pre-training data sets and the aspect of unfairness, that should be fixed. Biometric data should be administrated in the privacy and law, such as GDPR and EU AI Act.\u003c/p\u003e \u003cp\u003eLightweight deployment at the edge, federated training and privacy enforcement training, wide-range cross-environment generalization, continuous learning of dynamically changing databases, and explainable AI to understand errors with a clear vision are the areas that will allow making more progress in this direction in the future. Not only will high- tech expertise and ethical deployment and compliance be necessary to provide reliable, scalable, and fair automated attendance monitoring, but also ethical usage and enforcement.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe authors declare that no funding was received for this research.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eP.K. and R.R. conceptualized the study and designed the overall research framework. R.G. and M.K. conducted the literature review and analyzed existing deep learning models and facial recognition techniques. S. contributed to the methodology design, including face detection, feature extraction, and system pipeline development. A.K. assisted in data interpretation, comparative analysis, and preparation of tables and performance evaluation.P.K. and R.R. drafted the main manuscript text. R.G. and M.K. contributed to writing specific sections, including literature review and results analysis. S. reviewed the technical accuracy and refined the methodology section. A.K. prepared figures, formatting, and assisted in final editing.All authors reviewed, revised, and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eThalor MA, Gaikwad OS (2023) Facial Recognition Attendance Monitoring System using Deep Learning Techniques, Int. J. Integrated Sci. Technol., vol. 1, no. 6, pp. 841\u0026ndash;848, Dec\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFadel NEL (2025) Facial Recognition Algorithms: A Systematic Literature Review, J. Imaging, vol. 11, no. 2, p. 58, Feb\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaranya S, Chandru M, Ragul S, Kafeel SA (2024) Facial Recognition Attendance Monitoring System Using Deep Learning with YOLOv8. Int J Adv Res Sci Eng Manag (IJARESM)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAidana A et al (2025) Robust Face Recognition Under Challenging Conditions: A Comprehensive Review of Deep Learning Methods, Appl. Sci., vol. 15, no. 17, p. 9390, Aug\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Wu L, Zhou X, Fu H (2024) Beyond Surveillance: Privacy, Ethics, and Regulations in Face Recognition Technology, Front. Big Data, vol. 7, p. 1337465, Jul\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePainuly K, Bisht Y, Vaidya H, Kapruwan A, Gupta R (2024) Efficient Real-Time Face Recognition-Based Attendance System with Deep Learning Algorithms, in Proc. 2024 4th Int. Conf. Adv. Comput. Commun. Inform. Technol. (ACCIT), IEEE\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerma A, Gupta R, Sharma S (2023) Automated Attendance Monitoring System Using Facial Recognition, in Proc. 2023 Int. Conf. Emerging Trends Comput. Sci. Inf. Technol\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaq MU, Sethi MAJ, Ahmad S et al (2025) A Comprehensive Review of Face Detection/Recognition Algorithms and Competitive Datasets. Comput Mater Continua 84(1):1\u0026ndash;24\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJuwanda RC, Alunjati FA, Elviani U, Hidayat FP (2024) Comparative Analysis of FaceNet and ArcFace in Minimizing False Positives for Enhanced Access Control Security, in Proc. 2024 Int. Conf. ICT Smart Soc. (ICISS), IEEE, Sep. pp. 1\u0026ndash;6\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMustafa S, Wang J, Liu H (Jun. 2024) Towards Efficient and Robust Face Recognition Through Attention-Integrated Multi-Level CNN, Multimedia Tools Appl. Springer Nature\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaramizadeh S, Chaeikar SS, Salarian H (2025) Combining MTCNN and Enhanced FaceNet with Adaptive Feature Fusion for Robust Face Recognition, Technologies, Oct\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChansaeng T, Jarupunphol P (2024) Attendance System Optimization through Deep Learning Face Recognition, Int. J. Comput. Digit. Syst., vol. 15, no. 1, pp. 1527\u0026ndash;1540, Apr\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDang H (2023) Int Efficient Deep Learning Approach for Facial Recognition Using Improved FaceNet Based on MobileNetV2, in Proc. Conf. Comput. Sci. Inf. Technol., 2023\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJarupunphol P, Chansaeng T (2024) An Automated Face Detection and Recognition for Class. Attendance Using IoT and Edge Computing,\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernando R, Athauda H (2024) Image Processing Based Real-Time Online Attendance Monitoring System Using Facial Recognition, in Proc. 2024 Int. Conf. Image Process. Robot. (ICIPRoB), IEEE\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChin KC et al (2024) Deep Learning Based Attendance Check System, in Proc. 2024 9th Int. Conf. Intell. Inf. Technol. (ICIIT), ACM\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBalakannan SP, Sravani K, Jeyapriya S, Fathima, Rosini AIB (2024) Attendance System Using Machine Learning, in Proc. 2024 Int. Conf. Electr. Electron. Comput. Technol. (ICEECT), IEEE\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTalukder OM, Dakhil A, Sheikh M (2024) Automation of Surveillance Systems Using Deep Learning and Facial Recognition, J. Inf. Syst. Eng. Manag., vol. 10, no. 35s\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePonnusamy M et al (2024) Enhancing Security in Public Spaces Through Generative Adversarial Networks (GANs), J. Inf. Syst. Eng. Manag., vol. 10, no. 35s\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSadak MS (2025) Face Recognition Technologies: A Comprehensive Review. Int J Sci Basic Appl Res (IJSBAR) 77(1):333\u0026ndash;345\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbraham CS, Dev D, Patnaik R (2023) Int Challenges in Facial Feature- Based Recognition: Glasses, Beards, and Demographic Variability, in Proc. Conf. Comput. Commun. Inform., 2023\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBose S, Kumar M, Patel R (2024) Efficient Deep Models for Real-Time Face Recognition. IEEE Trans Inf Forensics Secur\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnonymous (2025) Evaluation of Deep Learning Methods in Face Recognition. Int J Sci Adv Technol (IJSAT)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaj B, Kumar S, Mehra P (2023) Comparative Analysis of Eigenfaces, Fisherfaces, and LBPH Under Variable Illumination for Attendance Systems. Int J Comput Appl 185(42):1\u0026ndash;7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnonymous AI-P (2025) Facial Recognition Attendance System, Int. J. Res. Sci. Innovation (IJRSI), vol. 12, no. 9, pp. 3902\u0026ndash;3912, Sep\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFadel N et al (2025) Facial Recognition Algorithms: Systematic Literature Review Focusing on Occlusion and Bias. J Imaging, 11, 2\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnonymous (2025) 50 Years of Automated Face Recognition, arXiv preprint arXiv:2505.24247\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnonymous (Jun. 2025) Global Perspectives on Regulating Facial Recognition Technology Utilization for Criminal Justice, Global Public Policy Governance. Springer Nature\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAidana A et al (2025) Deep Learning and Facial Biometrics: Compliance with EU AI Act Requirements. Appl Sci\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerma A, Kumar S (2024) Fairness and Bias Evaluation in Institutional Facial Recognition Attendance Systems, in Proc. 2024 IEEE Int. Conf. Biometrics Theory Appl. Syst. (BTAS), IEEE\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Facial recognition, deep learning, attendance monitoring, ArcFace, FaceNet, MTCNN, YOLOv8, convolutional neural networks, transformer models, biometric systems, edge deployment","lastPublishedDoi":"10.21203/rs.3.rs-9504761/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9504761/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe challenge for schools and organizations in need of an intelligent, automated, and real-time attendance monitoring system has become complex. In this survey, we present a survey of facial recognition frameworks based on deep learning and addressed the architectural patterns of CNNs (CNN), ResNets, FaceNet, ArcFace, CosFace, and Vision Transformers as well as face detection pipelines based on MTCNN, YOLOv5, and YOLOv8. We discuss the steps of a face detection pipeline\u0026mdash;face detection, preprocessing, feature extraction, matching, and database integration, and practical deployment scenarios. Our study demonstrates that deep learning models such as ArcFace and transformer models provide high accuracy for facial identity verification. Transfer learning techniques have effectively removed the need for large labeled institutional datasets while light versions of FaceNet and MobileNet-based models allow real- time processing on edge devices. Further, multi-face simultaneous recognition models with anti-spoofing modules provide more reliability for large- scale attendance applications. But, several obstacles are preventing widespread adoption. Performance under occlusion and varying illumination, demographic bias, privacy and regulatory concerns, and computational limitations still limit the adoption of attendance monitoring. Future research should focus on robust cross-environment models, fairness- aware training, edge-optimized architectures, and privacy- preserving models to ensure responsible and scalable attendance monitoring.\u003c/p\u003e","manuscriptTitle":"A Comprehensive Review of Deep Learning- Driven Facial Recognition Frameworks for Intelligent and Fully Automated Attendance Monitoring Systems","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 06:46:33","doi":"10.21203/rs.3.rs-9504761/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0a02cab6-f2ed-4f1c-a8e4-aa6096dcba17","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-03T04:12:25+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-01T13:24:03+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-03T04:24:45+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 06:46:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9504761","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9504761","identity":"rs-9504761","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}