Authors - Aneesah Sabar, KA Dilini T Kulawansa Abstract - Federated Learning has emerged as a robust privacy-preserving framework that enables joint model training across multiple distributed clients without sharing raw data. However, the effectiveness of traditional federated learning frameworks is hindered by client heterogeneity, where participants differ in data distribution, computational resources, and communication capabilities. This survey investigates evolution of personalization techniques in Federated Learning that address these challenges by tailoring models to individual clients while maintaining the benefits of global collaboration. The paper categorizes existing personalization approaches into five major groups: local fine-tuning, model interpolation, meta-learning methods, clustered federated learning, and regularization-based techniques. Each method’s core idea, strengths, limitations, and suitability under different heterogeneity conditions are analyzed in detail. The findings indicate that personalization significantly improves fairness, accuracy, and adaptability across heterogeneous clients, though it introduces trade-offs in communication cost, scalability, and privacy. This review concludes that personalization is essential for deploying federated learning in realistic, diverse environments and highlights emerging directions in fairness-aware, resource-efficient, and privacy-preserving personalization. Future research should focus on scalable and dynamic personalization strategies capable of handling evolving client behaviors and large-scale federated systems.
Authors - Bhavanam Sruthi, Mettu Sai Preethi, Krishna Reddy Abstract - Driver drowsiness is a major reason for accidents on the road, hence it is important to detect it early to increase safety on the road. A driver drowsiness detection system based on deep learning algorithms is proposed and it uses images captured through a camera installed inside a car. Various deep learning algorithms, namely CNN, VGG16, DenseNet121, MobileNet, LeNet, AlexNet, RNN, patchTST,Vision Transformer and Swin Transformer are implemented and compared to assess their performance.The system detects the conditions of the driver, whether eyes are open, closed, yawning, or not yawning. Among all these algorithms, the highest accuracy of 97.61% was obtained by using the MobileNet model, which proves that deep learning can play a vital role in detecting drowsiness. In addition, an alert can also be sent to warn the driver.
Authors - Prachita Chaudhari, Shubham Kishor Kadam, Shiwani wagh, Pankajkumar Anawade, Deepak Sharma, Chhitij Raj Abstract - One of such industries is healthcare, where data-driven methods attract much attention, and one among them is the field of predictive analytics that is already making a great difference in the healthcare industry concerning its capacity to enhance early diagnosis and treatment. Through this, whole care is provided, and this implies that the problem of fragmented care addresses systemic issues such as inefficiencies and inflation of costs. Predictive analytics is the most effective in the prediction of risks of disease, i.e. making care a thing related to the individual patient. Besides, it can be utilized to monitor populations and enhance the management of population health utilizing its combination of machine learning, natural language processing and deep learning. Solutions that offer the following benefits, including reduced misdiagnosis, re-source utilization, and affordable access to health care, are also being created with the help of the main enabling technologies including AI, IoT, and big data. Nevertheless, issues like the quality of the data, technical issues, ethical concerns and compliance with laws persist. Future work still to be done areas can be seen in the application of new technologies like quantum computing to answer questions about the public health, real time data that uses IoT, and the application of other mediating technologies in underserved locations that can instill equity and sustainability. Being fueled by the collaboration of various professionals, e.g., clinicians, data scientists, and policy-makers, predictive analytics is bound to enhance patient outcomes and catalyze the better provision of preventive, personalized, and responsibility healthcare solutions.
Authors - Binh Pham Nguyen Thanh, Chau M. Truong, Nhan Thi Cao Abstract - ResNet is widely used in medical image classification due to its strong hierarchical feature extraction capability. This study investigates the integration of Kolmogorov–Arnold Networks (KAN) and ConvKAN into ResNet to analyze the effect of increasing nonlinearity at different stages within a hierarchical skin lesion classification framework. Convolutional KAN is applied at the initial layer to enhance low-level feature extraction, while KAN is introduced at the final layer to improve high-level decision boundary modeling. A combined configuration is also evaluated to examine potential complementary effects across different levels of label granularity. Results show that performance depends on both the integration stage and dataset characteristics. Convolutional KAN at early layers provides limited and inconsistent improvements, whereas KAN at the final layer yields more stable gains. In addition, models incorporating KAN-based architectures generally achieve better performance across metrics such as accuracy, precision, F1-score, and ROC AUC. As classification becomes more fine-grained, Recall consistently decreases despite high ROC AUC, indicating challenges in decision thresholding across hierarchical levels. Overall, KAN is more effective for high-level decision making, while dataset complexity has a greater impact than architectural modifications.
Authors - Mouna Meghana Nagala, Anjan Babu G Abstract - Social Anxiety Disorder (SAD) remains one of the most pervasive mental health challenges globally, characterized by a debilitating “perception gap” where individuals consistently overestimate the visibility of their internal distress while underestimating their social performance. This paper introduces an Explainable AI (XAI) multi-modal sensing system designed for automated social anxiety monitoring and self-perception recalibration. The architecture is founded on an event-driven framework integrating real-time threedimensional facial feature encoding (DeepFace), acoustic prosody extraction (Librosa), and Natural Language Processing (NLP) for cognitive distortion detection. The system implements a Cognitive Behavioral Therapy (CBT) logic layer that provides interpretable feedback on linguistic patterns. System performance was benchmarked against the FER-2013 and RAVDESS repositories, yielding an anxiety detection sensitivity of 92.4% and a specificity of 94.7%. The findings affirm that coupling volumetric affective computing with generative AI constitutes a viable pathway toward trustworthy computer-aided detection (CAD) in behavioral health screening programs.
Authors - Theresa T. Limos, Sheena Sapuay-Guillen Abstract - This study developed PU-Serv: A Tool in Analyzing Student Services Using Machine Learning, a web-based system designed to enhance the evaluation of student services through automated sentiment analysis. The study assessed the existing student services evaluation form in terms of adequacy, efficiency, and reliability and aimed to develop a machine learning–based model to support the analysis of student feedback.A descriptive and developmental research design guided by Agile methodology and the CRISP-DM framework was employed. Data were gathered from focus group discussions, questionnaires, and institutional student feedback records. Natural language processing techniques were used to preprocess narrative feedback, and the Support Vector Machine (SVM) algorithm was integrated into the system due to its high accuracy in sentiment classification. The developed PU-Serv system automatically analyzes student feedback and presents summarized results through a web-based dashboard. The system provides administrators with actionable insights that support data-driven decision-making, helping institutions identify service issues, improve responsiveness, and enhance the overall quality of student services.
Authors - Neel Lathiya, Akshita Kadam, Amit Thakkar Abstract - Industrial tracking tools have led to the development of Quick Response codes, which are an essential component of digital engagement and provide simple access to payments, authentication, and online services with a single scan. However, they are very vulnerable to exploitation, particularly zero-click attacks, which start destructive operations without the user’s consent, due to their architecture, which is based on visual legitimacy, automatic intent execution, and plaintext encoding. This survey looks at the technical aspects of making and reading QR codes, charts the evolution of threats based on QR codes, ranging from physical manipulation to silent deep link hijacking, and explains how these attacks go beyond the robust security models of iOS and Android by utilizing trusted system paths. Based on five significant studies, we analyze real-world attack scenarios, user behavior gaps, and the efficacy of novel defenses like scanner assessment frameworks, zero-trust architecture approaches, and AI-driven payload inspection (AP3X, QRShield). Certain recommendations are made regarding system hardening, cryptographic integration, and user awareness in order to transform QR codes from a latent risk into a safe and verifiable medium.
