Univrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Adaptive Fingerprint Verification Using Siamese Neural Networks and Transfer Learning for Robust Authentication of Damaged Prints11610625110.22059/jitm.2026.106251ENSumanMAssistant Professor, Department of Artificial Intelligence and Data Science, B.M.S College of Engineering, Affiliated to Visvesvaraya Technological University, Belagavi, India.ShobhaNAssociate Professor, Department of Computer Science and Design, Dayananda Sagar College of Engineering, Bangalore, Visvesvaraya Technological University, Belagavi, India.SrideviMuruhanProfessor, Department of Biotechnology, Vinayaka Mission`s Kirupananda Variyar Engineering College, Salem (Vinayaka Mission`s Research Foundation), India.VinothkumarSAssistant Professor, (SRG), Department of Information Technology, Kongu Engineering College, Perundurai, Erode, India.RamkumarRAssociate Professor, Electronics and Communication Engineering, Bannari Amman Institute of Technology, Sathyamangalam.R JPoovaraghanDepartment of Computer Science and Engineering, Vel Tech Rangarajan Dr.Sagunthala R&D Institute of Science and Technology, Avadi, Chennai, India.Journal Article20260306Fingerprint recognition is a vital component of biometric authentication, yet its reliability often declines with damaged or partial inputs. This study introduces a fingerprint verification framework based on Siamese Neural Networks (SNN) with transfer learning to enhance adaptability and accuracy. Using the Sokoto Coventry Fingerprint Dataset (SOCOFing), fingerprint images were preprocessed into labeled pairs of similar and dissimilar samples for model training and evaluation. A distinctive feature of the proposed system is dynamic class expansion, enabling new user integration without retraining. The framework also ensures privacy-preserving verification by securely encrypting and storing extracted feature embeddings. Experimental results across four difficulty levels—real, easy, medium, and hard—demonstrate high reliability and robustness in both intact and degraded fingerprint conditions. Overall, the proposed approach offers a scalable, secure, and efficient solution for fingerprint-based identification, addressing key challenges in real-world biometric authentication systems.https://jitm.ut.ac.ir/article_106251_4d047dff1d53f3521454e4cdfaaca1e8.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Data-Efficient Transformer Architectures for Image-Level Facial Forgery Detection: A Comparative Evaluation of ViT and DeiT173310625210.22059/jitm.2026.106252ENAkshathaGAssistant Professor, Department of Computer Science& Engineering, B.M.S College of Engineering, Affiliated to Visvesvaraya Technological University, Belagavi, India.KempannaMAssociate Professor, Department of Artificial Intelligence and Machine Learning, Bangalore Institute of Technology, Visvesvaraya Technological University, Bangalore, India.Preethi KolluruRamanaiahCloud Architect, Lead of AI initiative Program, Ernst & Young LLP, New York, USA.BandiDossDepartment of ECE, CMR Technical Campus, Hyderabad, Telangana, India.RamaniSProfessor, ECE Department, Sreenidhi Institute of Science and Technology, India.KirubakaranRangasamyAssistant Professor, Department of Biotechnology, Vinayaka Mission`s Kirupananda Variyar Engineering College, Salem (Vinayaka Mission`s Research Foundation). India.Journal Article20260306The rapid development of deepfake technologies has increased the demand for a credible and inter-pretable system for facial forgery detection. This study compares two transformer-based architec-tures—Vision Transformer (ViT) and Distilled Data-Efficient Image Transformer (DeiT)—for de-tecting real and manipulated facial images. The study aims to measure performance in terms of de-tection as well as interpretability and to address the weaknesses of traditional convolutional models. Data augmentation was applied, and a balanced dataset containing 8,000 real and fake images was constructed; both models were then fine-tuned under the same training environment. The explanatory ability of the models was incorporated using LIME. Experimental findings indicate that both models perform well, with DeiT being slightly more accurate at 94.62% than ViT at 93.6%, alongside faster convergence rates and less overfitting. Visualization of the focus on important facial areas confirms that the models reliably register synthetic artifacts. Although promising, generalization across dif-ferent datasets and enhancement of real-time performance remain challenges. Overall, the results validate transformer architectures—especially DeiT—as powerful and explainable deepfake detec-tion algorithms, valuable for ensuring safe and transparent digital media forensics.https://jitm.ut.ac.ir/article_106252_3a0121a057e9925838edec110c5de17d.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Ball Screw Surface Defect Impact Level Categorization Using the Collective Constraint Intelligent Prediction Technique345210625310.22059/jitm.2026.106253ENSooryaPrabha MohanResearch Scholar, Department of Computer Science and Engineering, National Institute of Technology, Trichy, Tamil Nadu, India.S JayaNirmalaAssociate Professor, Department of Computer Science and Engineering, National Institute of Technology, Trichy, Tamil Nadu, India.Journal Article20260306Ball screw failures often exhibit early signs of surface defects, which can arise from contamination and indicate deterioration. Identifying these surface defects is crucial for minimizing repair costs and maximizing machinery uptime. This study presents a prediction algorithm based on transfer learning methods to enhance the accuracy of ball screw surface defect detection throughout its lifecycle. We investigate four transfer learning models: CCIP-V, CCIP-D, VGG16, and DenseNet, utilizing image data mining techniques for defect classification. These models are validated using a specialized dataset of ball screw surface defects, employing Region of Interest (ROI) masking techniques to enhance image classification for each model. Our findings reveal that the proposed hybrid approach, combining CCIP and ROI, demonstrates superior performance, with the best classifier achieving an accuracy of 0.983. Notably, the CCIP classifier, when enhanced with ROI techniques, achieves an impressive accuracy of 0.985, effectively predicting defect impact levels. This research underscores the potential of transfer learning and the integration of CCIP and ROI in improving classifier performance and identifying surface defect severity in ball screws.https://jitm.ut.ac.ir/article_106253_a256e9d74b9398b909f6e5783bbddcbf.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Adaptive Machine Learning Framework for Deepfake Detection: An Ensemble-Driven Approach with Optimized Feature Engineering537910625410.22059/jitm.2026.106254ENShruthiSResearch Scholar, Computer Science Engineering, R R Institute of Technology, Visvesvaraya Technological University, Belagavi, India.ManjunathRProfessor and Head, Computer Science Engineering, R R Institute of Technology, Bangalore, Visvesvaraya Technological University, Belagavi, India.Journal Article20260306With the rapid progress of artificial intelligence, it's now possible to generate extremely undoubted DeepFake images and videos, leading to major concerns related to misinformation, identity theft, and online security. Detecting such manipulated media requires robust and efficient techniques. Deep learning models, particularly Convolutional Neural Networks (CNNs), have been confirmed to be highly effective, but they typically demand high computational resources and may suffer from overfitting issues. "This study presents a Machine Learning based approach for the detection of DeepFakes, utilizing Histogram of Oriented Gradients (HOG), utilizing Principal Component Analysis (PCA) to decrease the dimensionality, and employing a combination of Random Forest and XGBoost classifiers for the final prediction. The DeepFake Detection Challenge Dataset, sourced from Kaggle, is employed for both training and performance assessment. The methodology consists of feature extraction using HOG to capture texture and gradient information, dimensionality reduction with PCA to enhance computational efficiency, and ensemble classification with Random Forest and XGBoost to improve detection accuracy. Hyperparameter tuning is performed to further optimize model performance. The investigational results prove that the anticipated hybrid ML model accomplishes an accuracy of 95.5%, outperforming conventional standalone classifiers and providing a computationally efficient alternative to deep learning-based approaches. This study highlights the efficiency of merging traditional feature-engineering techniques with ensemble learning for DeepFake detection. The results contribute to ongoing research in AI-driven security and media forensics, offering a scalable, interpretable, and high-performance solution for identifying manipulated media. Future work can explore additional feature extraction techniques and ensemble models to further enhance detection capabilities.https://jitm.ut.ac.ir/article_106254_b898ca3ec9b7eb14f35c304df1bccabd.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Secure Medical Data Transmission Using a Deep Learning–Based Quantum-Resistant Hybrid Stegno-Crypto Model8010110625510.22059/jitm.2026.106255ENNikhilaSResearch Scholar (Part Time), Research Centre -Department of Electronics and Instrumentation Engineering, Dayananda Sagar College of Engineering, Bengaluru, Karnataka, India; Assistant Professor, Department of Electronics and Instrumentation Engineering, Dr. Ambedkar Institute of Technology, Bengaluru, Karnataka, India.KrushnasamyV SAssociate Professor, Department of Electronics and Instrumentation Engineering, Dayananda Sagar College of Engineering, Bengaluru, Karnataka – 560078 Affiliated to Visvesvaraya Technological University, Belagavi, Karnataka, India.Journal Article20260306This study focuses on the secure transmission of medical data, an important element of contemporary healthcare systems in which sensitive patient data is shared over digital networks to facilitate diagnosis, monitoring, and decision-making. Due to the growing volume of medical records and electronic medical imaging data, it has become critical to ensure the integrity, confidentiality, and accuracy of the data throughout transmission. The two significant problems the study will cover include high-frequency noise in medical records that reduces diagnostic accuracy and the high sensitivity of traditional encryption techniques to quantum-computing-based attacks that jeopardize data privacy. To address these problems, a smart, noise-hardy, and quantum-resistant dual-layer design is proposed to provide an efficient and secure transfer of medical information without compromising quality or computational efficiency. The methodology includes two major steps. In the first step, a Noise-Resilient Pre-Encryption Data Pre-processing Algorithm is created using a Random Forest-Wavelet Filter hybrid model, in which the Random Forest employs adaptive parameter selection and the wavelet filter employs multilevel decomposition to remove noise. Genetic algorithms (GA) are used to optimize these parameters more dynamically in response to shifting noise levels. PSNR and SSIM are used to validate the pre-processing performance, and it is proven that the data fidelity is enhanced. The second phase proposes a Hybrid Quantum-resistant Steganography-Cryptography Algorithm that combines CNN-based Steganography with AES encryption. The CNN safely encodes medical information in non-sensitive carriers using encrypted messages to transmit it invisibly without tampering. In contrast, the AES uses dynamic key generation to resist classical and quantum attacks. The results of the experiment show that accuracy (96.89%), precision (89%), recall (94.50%), F1-score (91.60%), and system efficiency were increased significantly, with computational time (8.47 ms), latency (8.50 ms), and throughput (117.60 OPS) also increasing.https://jitm.ut.ac.ir/article_106255_238867e26e3771e837e69625c09ccc62.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301An Intelligent Hybrid Feature-Engineering Approach for Covariant Face Recognition Using Deep Learning10212210625610.22059/jitm.2026.106256ENPavithaU SResearch Scholar (Part Time), Research Centre - Department of Electronics and Communication Engineering, M. S. Ramaiah Institute of Technology, Bengaluru, India; Assistant Professor, Department of Electronics and Communication Engineering - M. S. Ramaiah Institute of Technology, Bengaluru, India.SumaK VAssociate Professor, Department of Electronics and Communication Engineering, M. S. Ramaiah Institute of Technology, Bengaluru, India; Affiliated to Visvesvaraya Technological University, Belagavi-590018, Karnataka, India.Journal Article20260306Face recognition (FR) is a non-contact biometric method integral to national and social security. It is pivotal in sectors like security, healthcare, banking, and criminal identification. Various techniques are being developed, including appearance and hybrid approaches, which either target specific facial features or consider the whole face for effective image recognition. This study explores hybrid machine learning techniques and enhanced covariates related to FR. The suggested approach is examined from a number of input viewpoints, including illumination, position variation, facial emotions, occlusions, and aging, which resulted in the widespread use of FR systems. The Generative Adversarial Network (GAN) is used to multiply the image on the dataset for image augmentation purposes. The facial covariants are extracted with the help of a hybrid feature engineering technique, such as a voting classifier that includes algorithms like K Nearest Neighbour, Support Vector Machine, and Random Forest Algorithm (KNN-SVM-RF). The dimension redundancy is achieved with the help of a combination of Principal Component Analysis and Independent Component Analysis algorithms. The modified VGG-16 algorithm is used to predict the image with the covariant similarity percentage of a person. Experiments conducted on the CelebFaces Attributes (CelebA) Dataset and Celebrity Face Image Dataset demonstrate that the hybrid strategy yields superior accuracy and robustness compared with CNN-only or classical approaches, achieving notable improvements in Precision, Recall, F1 score, and overall Accuracy.https://jitm.ut.ac.ir/article_106256_349d9d29e312c12e24179176a5f899e2.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Attentional Deep Learning with Inverse Transform Sampling for Robust Respiratory Sound Classification12314010625710.22059/jitm.2026.106257ENHemanthK SDepartment of Computer Science, Christ University, Bengaluru, India.HarishaNaik TDepartment of Computer Applications, Presidency College, Bangalore, India.NKartikDepartment of Commerce, Manipal Academy of Higher Education, Manipal, India.KumarN NandaAssistant Professor, Department of Computer Science and Engineering, Excel Engineering College Komarpalayampalyam, Tamilnadu, India.SenthilkumarSAssociate Professor, Department of Computer Science and Engineering, Vinayaka Mission`s Kirupananda Variyar Engineering College, Salem (Vinayaka Mission`s Research Foundation), India.RamyaRAssociate professor, Department of Computer Science and Engineering, Bannari Amman Institute of Technology, Erode, Tamilnadu, India.Journal Article20260306The necessity for efficient breathing sound classification systems originates from respiratory diseases, which impair oxygen-carbon dioxide exchange and impact lung function. Feature extraction and pattern categorization are general components of such systems. Because of their effectiveness with big datasets, deep neural networks have acquired popularity recently in the category of breathing sounds. Enhancing medical care requires cooperation amongst researchers, medical professionals, and patients. An attentional deep learning model with inverse transform sampling is presented in this study to classify respiratory diseases from audio data. Robust models were developed to classify and detect respiratory elements using the Respiratory Sound dataset. The primary objectives include effectively determining lung sounds and determining respiratory illnesses. The architectures of CNN, VGG16, and ResNet50 were developed to extract features and categorize data. Also, the pre-trained models ResNet50 and VGG16 identify critical characteristics in spectrum pictures more accurately. Inverse transfer sampling is used to rectify class imbalance in respiratory datasets. The models achieved 98% accuracy with the CNN model, 83% accuracy with VGG16, and 95% accuracy with ResNet50. Moreover, LSTM and CRNN models offer more information on how respiratory illnesses are classified.https://jitm.ut.ac.ir/article_106257_af73af2925ce5395caff52ce1e45f8d1.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301Hybrid Deep Learning Model for IVF Outcome Prediction from Time-Series Hormonal Data14115710625810.22059/jitm.2026.106258ENShaliniB NResearch scholar, School of Computer Science and Engineering, REVA University, Bengaluru, Karnataka, India.LithinKumbleAssociate Professor, School of Computer Science and Engineering, REVA University, Bengaluru, Karnataka, India.Journal Article20260306Optimizing results in assisted reproduction requires tailoring the dosage of follicle-stimulating hormone (FSH) during controlled ovarian stimulation (COS), but this is still challenging because of patient variability. Real-time modifications during stimulation are not supported by the majority of current models, which are restricted to static forecasts of starting dosages. This study proposes an advanced AI-driven framework for forecasting hormone dynamics and improving reproductive outcome prediction in IVF cycles. The methodology integrates multi-source clinical data with high-resolution time-series hormone profiles collected from Day 1 to Day 30 of ovarian stimulation. Data pre-processing includes normalization of hormonal values, alignment of temporal and static clinical attributes, and creation of patient-level merged datasets. Feature engineering incorporates daily hormone variations, moving averages, peak detection, and wavelet-based temporal pattern extraction, alongside encoded and normalized clinical parameters. For hormone trend forecasting, a hybrid deep learning model is developed that combines Wavelet Transform for noise reduction with LSTM and Transformer layers for sequential representation learning. The architecture captures short-term hormone fluctuations and long-range temporal dependencies, enabling accurate next-day hormone prediction. Model performance is optimized using RMSE = 0.31, MAE = 0.22, and MAPE loss metrics. This integrated approach enhances predictive accuracy = 88.9%, facilitates early-cycle monitoring, and supports clinical decision-making in assisted reproductive treatments.https://jitm.ut.ac.ir/article_106258_a3045edf00e55c972c4b03dfe26aa3bb.pdfUnivrsity Of Tehran PressJournal of Information Technology Management2980-797218120260301EduVeriChain: A Blockchain-Based System for Secure Academic Credential Verification and Management15818310653610.22059/jitm.2026.106536ENSheelaD VResearch Scholar, School of Science & Computer Studies, CMR University, Bengaluru, Karnataka, India.Ashok KumarT AProfessor and Director, School of Science & Computer Studies, CMR University, Bengaluru, Karnataka, India.Journal Article20260413Academic credential fraud is a pervasive issue with significant implications for educational institutions, employers, and society as a whole. The centralized nature of traditional academic verification systems makes them vulnerable to inefficiencies, corruption, and forgery, highlighting the need for a more secure and transparent solution. Blockchain technology provides a decentralized, tamper-proof framework for managing digital credentials, ensuring authenticity and integrity without intermediaries. This research proposes EduVeriChain, a blockchain-based system for securely issuing, verifying, and managing academic credentials. Utilizing smart contracts, decentralized storage via the InterPlanetary File System (IPFS), and robust encryption, EduVeriChain ensures the transparency, security, and immutability of academic records. The system’s Credential Revocation Authority (CRA) and Credential Status Enforcer (CSE) provide automated and secure revocation, further enhancing trust. Performance evaluation reveals that EduVeriChain achieves high throughput, low latency, and predictable costs, making it a scalable and practical solution for academic credential verification. By providing a decentralized, secure, and efficient framework, EduVeriChain aims to mitigate pervasive document fraud and promote integrity in academic credential management.https://jitm.ut.ac.ir/article_106536_e05448c07f334f89a9fd1bebcfde5ceb.pdf