ABSTRACT:
The growth of online payment industries and Online businesses have brought forth some significant security challenges, especially when it comes to credit card use that is unauthorized. In this paper, an innovative framework that detects fraud in real-time has been presented, immediately alerting the stakeholders of suspicious to an activity. We compare several classification methods with real trans- action data, with the performance of the methods measured in terms of several important performance measures such as accuracy, precision, recall and F1-score. We find that the methods of ensemble learning provide this benefit over both standard algorithms, namely better accuracy and reduced levels of false positives. The combined warning system gives the user the opportunity to respond instantly to protect themselves. Our study shows that smart analytical systems can be used to enhance security procedures in modern financial ecosystems.
Keywords: Credit card fraud, machine learning, ensemble methods, real-time detection, alert system, imbalanced dataset.

ABSTRACT:
This paper presents a nano-power cascoded β-multiplier CMOS Voltage Reference without the use of passive filters. The circuit includes a start-up, output stage, β-multiplier circuit, bias current generator, and operational amplifier to achieve a high-power supply rejection ratio and low line sensitivity. The cascoded β-multiplier CMOS Voltage Reference is implemented on 180 nm of technology. The reference voltage is 430.23 mV operating from a wide supply range of 0.9 V to 3.0 V. The Power Supply Rejection Ratio is -90.86 dB at 1 kHz, and the line Sensitivity is 0.0008 %/V. Furthermore, the design achieves an ultra-diminished power consumption of only 379.16 nW, with a total output noise of 27.17 nV/√Hz and draws a mere 421.29nA of supply current at 0.9V. The measured TC was found to be 58.2 ppm/°C.
Keywords: Beta-Multiplier, Power Supply Rejection Ratio, Line Sensitivity, ultra-low power consumption, Temperature Coefficient

ABSTRACT:
In criminal investigation, today, digital images are significant; however, they are manipulable by using sophisticated editing software, and it is challenging to detect the manipulation. Traditional forensic methods mainly rely on hash algorithms to check for tampering. How-ever, these tools are unable to detect advanced tampering techniques such as copy-move attacks and metadata manipulation. This research proposes a hybrid model that combines different techniques to detect tampering, including cryptographic verification, error level analysis, metadata analysis, and deep learning classification. The system combines the results from all these methods using a weighted confidence scoring system, which calculates a final tampering probability score. Moreover, it includes an immutable Chain-of-Custody logging system to provide forensic traceability, addressing important gaps in modern judicial documentation requirements.
Keywords—Digital Forensics, Image Tampering, Law Enforcement, Deep Learning, Chain-of-Custody, Cryptography.

ABSTRACT:
Aspect- Based Sentiment Analysis (ABSA) is a crucial task in natural language processing that allows companies to derive fine-grained information about customer reviews. However, current approaches are often unable to capture the aspect-context interactions, model a dynamic sentiment class, and class imbalance challenges. Despite the promising outcomes of pre-trained transformer models like BERT, they do not incorporate explicit category-aware mechanisms or effective regularization strategies. This paper proposed a model to address these shortcomings with a new model, RoBERTa-CA-WMA (Category-Aware Weighted Multi-Head Attention), which is combines RoBERTa with a category-aware attention mechanism. The model introduces an aspect projections layer to improve the interaction between semantics of the aspect embedding and the contextual features, and advanced regularization methods, such as contrastive loss to better differentiate features, label smoothing to prevent the issue of overconfidence, and the use of multi-dropout to encourage generalization. Large-scale experiments are conducted on benchmark datasets such as SemEval-2014 (Laptop and Restaurant), SemEval-2015 (Restaurant), and SemEval-2016(Restaurant). The proposed model achieves strong performance, obtaining Accuracy scores of 90.68% (Lap14), 95.53% (Rest14), 87.50% (Rest15), and 97.08% (Rest16), along with corresponding F1-scores of 90.53% (Lap14), 95.47%(Rest14), 88.41% (Rest15), and 96.95%(Rest16), respectively. They indicate a dramatic improvement over state-of-the-art models, especially with respect to the neutral sentiment handling, as well as in the event of an imbalanced data distribution. The major contributions of this work include a category-aware attention mechanism and RoBERTa, the addition of an aspect projection layer to allow better semantic alignment, and the use of advanced regularization to enhance robustness and generalization. Overall, the proposed framework demonstrates that a reinforced integration of powerful pre-trained transformers and category-sensitive attention and effective regularization have a significant impact on the performance and reliability of the ABSA systems.
Keywords: Aspect-Based Sentiment Analysis, RoBERTa, Category-Aware Attention, Contrastive Loss, Customer Reviews, Transformer Models, Multi-Head Attention, SemEval Benchmark.

ABSTRACT:
The expansion of Internet of Things (IoT) has reshaped the daily life by connecting wide range of smart devices, that work intelligently and autonomously. The large volumes of data, requires high computational power and extensive storage, which is provided typically by cloud platforms. However, depending solely on Cloud for time-sensitive or mission critical applications can pose challenges, as they require minimal service latency and immediate responses. Thus, Edge computing has emerged as a layer that enables real time and localized processing of the tasks. With fluctuations in data load, the performance of edge nodes may vary depending on their resources and capacity. So, it is essential to schedule tasks based on the priority and available resources, which enhances the Quality of Service (QoS) and the performance of edge nodes. Queuing Theory, which is based on task arrival and service patterns, provides a structured framework for modelling such systems with either single server or multiple servers. The Queuing theory presents different models such as, single-server and multi-server models which can be employed to schedule tasks based on the available resources at the servers and improve overall performance of the system. This study provides an analytical approach by comparing these two models at the Edge Computing layer, which is based on time taken for executing all the data packets in the system, average waiting time or average length of the system.
Keywords: IOT, Edge Computing, Queuing Models, Load Balancing

ABSTRACT:
The creator economy, estimated to be worth more than 2.1 billion dollars and expanding rapidly, has enabled individuals to monetize their skills. However, one of the most critical challenges remains finding collaborators with complementary abilities and aligned creative goals. Reports indicate that approximately 71% of creators (around 50 million globally) struggle to discover suitable talent, which affects project ambition, efficiency, and long-term sustainability. To address this gap, we present Creatist, an iOS application developed using design thinking principles and focused on intelligent collaborator discovery and streamlined collabo-ration workflows. Beyond networking, the platform integrates intelligent search and filtering, portfolio discovery, project management, deadline-driven task control, centralized file sharing, and real-time communication via chat and calls. These capabilities support meaningful partnerships, improve execution quality, and increase visibility of both individual and team outcomes. By improving collaboration workflows, Creatist contributes to stronger creator productivity and economic opportunity, with prior studies showing collaborative projects can earn up to 30% more than solo efforts.
Keywords: Design thinking, creator economy, talent discovery, iOS application, collaboration platform, intelligent search, SDG 8.

ABSTRACT:
Regarding the classical FL problem, it is well-acknowledged that there needs to be common feature space among the clients. In reality, this idea does not work because clients have their own datasets in different formats and, therefore, have different feature spaces. Only recently, attempts were made to solve this issue by making sure that the representations of the clients match some global anchor distribution using Wasserstein distance. Still, such an approach is still based on some predetermined assumptions about the feature distributions and embeddings. The work presents an innovative approach that combines VAEs with contrastive learning in order to tackle the main limitations in federated learning, particularly those that occur in a non-IID and heterogeneous environment. Every client uses VAEs in order to construct a probabilistic representation of their own data, allowing for creating a shared latent space between all the clients. We do not compare features based on the global anchor, but rather utilize contrastive learning approaches like InfoNCE or Supervised Contrastive Loss. It avoids the requirement of distribution estimates for each class and improves the versatility and expressivity of the learned representation. The suggested model is suitable for federated settings in both supervised and unsupervised fashions, which allows collaborative personalization in an environment with low communication resources and heterogeneous clients. This thesis evaluates the proposed framework on synthetic as well as real-world benchmark datasets. From our results, it is evident that the approach we have proposed outperforms the approaches using Wasserstein distance in case of scarce communication resources and high variance of features.
Keywords: Federated Learning, Feature Space Heterogeneity, Personalized Federated Learning, Variational Autoencoder, Contrastive Learning.

ABSTRACT:
Diabetes mellitus (DM) is a widespread metabolic illness. The International Diabetes Federation has predicted that the disease will affect 700 million individuals globally by 2045. The clinical burden of the disease is growing at an unsustainable rate with multiple AI-based predictive models demonstrating clinical promise, but few are actually being implemented in clinical practice. Machine learning has changed the landscape of diabetes prediction, and it is now possible to analyse heterogeneous clinical datasets at a scale that was previously not possible. This review summarizes data on datasets, algorithms, training protocols, and measures of evaluation to characterize the structural constraints to clinical translation. A Gradient-boosted model, Extreme Gradient Boosting (XGBoost) and an Adaptive Boosting model (AdaBoost), were found to work better on structured tabular data on the Pima Indians Diabetes Database (PIDD) and the Behavioural Risk Factor Surveillance System (BRFSS), whereas transformer architectures and Convolutional Neural Networks (CNNs) have become essential for Continuous Glucose Monitoring (CGM) time-series analysis. Although the Sylhet EMR dataset was included only for context purposes and not evaluated with respect to the cross-dataset architectural conclusions. Taken together, these results affirm that the bottleneck in algorithmic performance is no longer the dominant one. The other obstacles are structural, and federated learning, structural causal modelling, and model distillation must take the forefront in the future to decrease the burden of this widespread disease in the world.
Keywords: Diabetes, Machine Learning, Explainable AI, Diabetes Prediction, Diabetes Management.

ABSTRACT:
Multiplication is a basic arithmetic operation used in various digital signal processing systems, image processing systems, cryptographic systems, and high-speed computation systems. The performance of the above systems is highly dependent on the propagation delay of the multiplier architectures. This paper proposes an optimized 16×16 multiplier architecture to enhance the computation speed. The computation speed of the proposed multiplier is improved by optimizing the partial product generation. The proposed multiplier architecture is developed by using a hierarchical modular structure and a column-wise accumulation method by using the Lilavati addition method. The proposed design is coded using Verilog HDL and implemented on a Xilinx Artix-7 FPGA platform by using the Vivado design tools. The result obtained by implementing the proposed multiplier architecture on the Xilinx platform shows a propagation delay of 20.81 ns. The maximum frequency of the proposed multiplier is approximately 48 MHz. The propagation delay of the traditional multiplier architectures proposed by various authors using the same implementation platform is within the range of 23 ns to 80 ns. The proposed method outperforms the traditional method by achieving a lower propagation delay.
Keywords: Vedic Multiplier, Lilavati Addition (LA), Urdhva-Tiryagbhyam (UT), Population Counter, 16-to-5 Compressor.

ABSTRACT:
In RPL (routing protocol for low power and lossy network, according to RFC 6206, The Trickle algorithm uses a redundancy constant (k) that discard control message transmissions which is not useful. standard RPL uses a static value of k (typically 10) which is unable to adapt the varying network densities, due to excessive control traffic overhead in dense networks and slow to response in sparse networks. This paper proposed Density-Aware Dynamic k model, that adjusts the value of Redundancy constant (K) according to the network conditions. Simulations conducted with 10 to 60 sensor nodes using Cooja (Contiki-3.0) simulator. The results show that, compared to standard RPL proposed approach improves packet delivery ratio, reduced energy consumption and enhances network life time.
Keywords: RPL, redundancy constant (K), Trickle Algorithm, IoT, Control traffic Overhead.

ABSTRACT:
Psoriasis is a chronic inflammatory disease that occurs due to the immune system and affects multiple organs of the body. Around 125 million people worldwide live with this disease. This disease not only affects multiple organs of the body, but also impacts time, financial costs, and mental health. Due to targeted biological treatments, the treatment landscape has also improved, but clinical management is still a challenge, mainly due to the shortage of dermatologists and the subjectivity of visual assessment methods. This review analyzes the massive shift from traditional machine learning to multimodal foundation models and deep learning architectures that enable the standardized and reproducible Psoriasis Area and Severity Index (PASI) scoring. We carefully analyze the role of molecular AI, specifically spatial transcriptomics(ST) and single-cell resolution, which helps in reversing the effect of convolution on localized inflammatory niches, and also the emerging “digital twins” for predictive modeling of treatment response. Additionally, we address the computational frameworks that enable algorithmic precision medicine, including generative data augmentation, privacy-preserving federated learning, and explainable AI(XAI) frameworks. In the end, we conclude that the successful clinical translation of these technologies requires a “human-in-the-loop” approach so that we can navigate through the social and regulatory challenges with ethical obligations.
Keywords: Psoriasis; Deep Learning; Multimodal Foundation Models; Psoriasis Area and Severity Index (PASI); Spatial Transcriptomics; Digital Twins; Federated Learning; Explainable AI (XAI).

ABSTRACT:
Short-term load forecasting (STLF) is still an essential operational necessity in today’s electric power system, especially in relation to the increasing variability in electric demand, renewable energy integration, and development of smart grid infrastructure. The application of statistical techniques and machine learning algorithms in STLF has been highly promising. However, most of the techniques used in STLF have been based on fixed model or ensemble strategies that do not account for input-dependent variations in electric demand. Such inflexibility in dealing with electric demand is likely to limit their effectiveness in dealing with non-linear, non-stationary, and peak demand conditions. To fill in this knowledge gap, this paper proposes an adaptive statistical modelling approach with multi-model switching for short-term electric load forecasting. The proposed approach combines seven state-of-the-art models, namely linear regression, ridge regression, Lasso regression, Support Vector Regression (SVR), Random Forest, CatBoost, and LightGBM. Furthermore, it proposes an adaptive switching approach that combines model predictions using instance-level model performance. The proposed approach is tested using hourly electric demand data from the New York Independent System Operator (NYISO) dataset. A comprehensive set of preprocessing techniques, including normalization, outlier handling, and feature development, is also used. The experimental results validate that the proposed approach performs better than existing techniques with an accuracy increase of 3-5%. Moreover, the proposed approach also shows improved performance in terms of reduced root mean squared error. Furthermore, statistical validation is also used to confirm the effectiveness of the proposed approach
Keywords: Adaptive Switching Mechanism, Autoregressive Integrated Moving Average (ARIMA), Categorical Boosting (CatBoost), Gradient Boosting Machine (GBM), Short-Term Load Forecasting (STLF)

ABSTRACT:
The presence of fog and haze significantly reduces the efficiency of visual perception algorithms for safety-critical applications, such as self-driving cars, surveillance systems, and aerial imaging. Existing lightweight dehazing models utilise a constant restoration power regardless of the level of foggy atmosphere, which results in systematic overdehazing for images of light haze and underrestoration for images with dense fog. We propose AdapHaze, an adaptive two-stage dehazing method that allows addressing this problem with density-based image restoration. In the first stage, a MobileNetV3-Small classifier trained on a subset of the RESIDE OTS dataset balances three categories of foggy atmosphere: light, medium, or dense, obtaining very high validation accuracy. In the second stage, a trained AOD-Net model uses a combined MSE and VGG-based perceptual loss function for image restoration with intensity set depending on the value of the density-conditioned lambda parameter obtained from the result of the first stage prediction. Evaluating the model on the benchmark test RESIDE SOTS-Outdoor dataset, we demonstrate competitive fidelity results compared to the baseline and the fixed-lambda perceptual model. The entire two-model pipeline has 1.65M parameters and executes in about 15ms/image, making it ideal for deployment in real time. The evaluation through density stratification reveals that there are performance gains in all density classes of fog, thus proving that the proposed adaptive λ technique is indeed an effective and efficient approach to dehazing images.
Keywords: Image dehazing, fog density estimation, MobileNetV3, AOD-Net, perceptual loss, adaptive processing, atmospheric scattering

ABSTRACT:
This review traces the evolution of the last two decades of computational forecasting models for Alzheimer’s disease progression, where the field has evolved from early linear methods to advanced multimodal and foundation models. Therefore, the paper highlights the computational importance of early disease prediction through longitudinal cohorts (e.g., ADNI, OASIS-3) and multimodal data architectures, which include structural, functional, clinical, biofluid, and genomic biomarkers. The review discusses the trajectory of predictive modelling from initial cognitive and structural pairing techniques to the deep learning era, where automated feature extraction and temporal tracking were introduced. Importantly, this review addresses a significant “Credibility gap” in recent deep learning research, which occurred due to data leakage and improper data-splitting techniques. So, the paper advocates a strict “Methodological Triad", in which subject-level data separation and external validation are essential. Furthermore, this review analyses the state-of-the-art paradigms of 2024-2026, where Transformer-based architectures, tabular foundation models, and Agentic AI frameworks are integrating, which are shifting the field from black-box prediction to explainable clinical intelligence. This review examines XAI techniques such as SHAP, Grad-CAM, LIME, and LRP. Along with this, Graph Neural Networks(GNNs) are discussed as a framework of a brain connectivity model, where cortical regions are nodes and their anatomical and functional connections are edges. Because of this, network-level disruptions can be captured that cannot be detected by scalar biomarkers.
Keywords: Alzheimer’s Disease Forecasting; Multimodal Deep Learning; Longitudinal Cohorts; Foundation Models; Data Leakage; Methodological Triad; Agentic AI; Neuroinformatics; Decentralized Learning; Explainable AI (XAI); Graph Neural Networks; Brain Connectivity.

ABSTRACT:
The early detection of crop stress is essential to increase agricultural output but much of the current solution involves using a multispectral sensor or satellite-based applications that are not available to small and medium-scale farmers. This article suggests the application of a cost effective web based stress prediction system in wheat, maize and rice using images that are taken using the traditional RGB cameras. The system does not need any specialised hardware, and it is based on the analysis of visible-spectrum images and some simplified user-provided agricultural data. Images of uploaded plants are processed on the pixel level before extracting the RGB values, which are used to approximate an NDVI to determine the health of crops. Other inputs such as the type of crop, month of sowing, and location identify the growth stages and access the current weather information through external APIs. The resulting values are fed into a rule-based decision mechanism to put the degree of crop stress in a category and provide simple actionable recommendations to farmers. The results are all stored and visualised using an interactive dashboard making historical analysis and tracking easier. Empirical evidence shows that the framework is an effective, cost-effective, and resource-accessible approach to crop health tracking through efficient data analysis and visualization of stresses to detect trends in crop health.
Keywords: crop stress recognition, NDVI estimation, image processing, smart agriculture, web-based system, crop health monitoring.

ABSTRACT:
In this paper, UWB band microstrip patch antenna has been designed to detect an early stage of breast cancer detection by using confocal microwaves imaging technique. A patch antenna should be fabricated over FR-4 substrate, with dimensions of 22x30 mm, having hexagonal circular patch with defected ground structure. The antenna has designed to achieved a moderate peak gain of 2.6 dB with average depth of -55 dB return loss at resonating frequency of 4.4 GHz. Antenna resonating at sharp 4.4 GHz with proper matching impedance with efficiency of 90% to penetrate the waves deep inside the breast to detect the correct location of cancer cell with high accuracy while maintaining low system cost and complexity. The UWB microstrip antenna are placed Infront of the breast model to transmit the EM waves and receive the scattered waves from tumor cell which having high dielectric property to absorbed the more energy than healthy cell. Finally, the scattered waves collected and proceed to Confocal Microwaves Imaging technique in MATLAB to regenerate the high resolution 2D image by using Delay-and-Sum Beamforming algorithm which represent the tumor position deep inside the breast. The regenerated image shows the high intensity of energy showing the more absorption of energy by tumor cell.
Keywords: Microstrip Patch Antenna, Breast phantom model, Confocal Microwave Imaging, MATLAB, -CST

ABSTRACT:
The COVID-19 pandemic is the catalyst that led to the greatest activity on social media, making Twitter an important source of information about the opinion of the masses, emotional reactions and the trends in behavior under the influence of the global crisis. With over 500 million tweets being posted every day, during the most critical stages of the pandemic, researchers had access to new corpuses of real-world sentiment information never before. Lexicon-based, machine learning, deep learning, and transformer-based methods have been applied in the past to analyze COVID- 19-related tweets in locations and languages. The paper provides a systematic review that synthesizes the data collection processes, pre-processing pipelines, sentiment classification processes, nation-specific sentiment trends, emotion and topic modelling insights, model performance comparisons, research limitations, and research gaps in Twitter-based COVID-19 sentiment analysis. Our review of over 30 peer-reviewed papers shows that transformer models, and more so BERT variants are never worse than traditional methods, and it has still much to offer in terms of multi-lingual analysis, real-time tracking, and epidemiological integration. The article is a well-structured information reservoir of the researchers and practitioners who intend to work on the domain of crisis-based social media analytics.

ABSTRACT:
ThisEarly detection of cardiovascular diseases is essential to reduce mortality and enable timely medical intervention. This paper proposes a real-time IoT-based heart disease prediction system that integrates multiple physiological sensors and machine learning algorithms for continuous health monitoring. The system incorporates a heartbeat sensor, Galvanic Skin Response (GSR) sensor, pulse oximeter for SpO₂ and pulse rate measurement, and a DS18B20 temperature sensor. Sensor data is acquired through the MCP3008 ADC and processed using a Raspberry Pi 4. A Random Forest machine learning model is employed to analyse physiological patterns and predict potential cardiac abnormalities. The system provides real-time visualization via an LCD display and sends alerts through a GSM module when abnormal conditions are detected. Experimental evaluation shows that the proposed model achieves an accuracy of 92.3%, precision of 91.1%, recall of 90.5%, and F1-score of 90.8%, demonstrating its effectiveness for early heart disease prediction and continuous monitoring.
Keywords: Heart Disease Detection, Machine Learning, IoT Healthcare, Real-Time Monitoring, Random Forest, Early Warning System.

ABSTRACT:
Deep learning models have achieved remarkable success in medical image analysis in recent times. However, their tendency to produce overconfident predictions even under high uncertainty, limits their reliability in safety-critical applications. In the clinical setting, minimizing the incorrect predictions is often more important than maximizing the overall coverage. This paper investigates the use of selective prediction via training dynamics (SPTD) to improve diagnostic reliability by enabling models to abstain from uncertain predictions. Unlike the conventional confidence based approaches, SPTD leverages prediction instability across training checkpoints as a proxy for uncertainty. Experiments are conducted on three different medical imaging datasets including diabetic retinopathy, urine cell classification and bone tumor classification datasets. Results clearly demonstrate that SPTD has consistently improve the prediction accuracy at lower coverage levels and achieved more than 20% improvement compared to non-SPTD model. The findings highlight the effectiveness of training dynamics as a robust uncertainty estimation mechanism and establish selective prediction as a viable strategy for reducing diagnostic errors in healthcare AI systems.
Keywords: Deep learning, Medical Imaging, Selective Prediction, Training Dynamics, Uncertainty Estimation, Abstention.

ABSTRACT:
Heart disease is the number one cause of death across the globe with millions of deaths annually and is a serious challenge to healthcare systems in the world today. The early and efficient diagnosis is vital in guaranteeing the timeliness of treatment and the management of the disease which may further decrease the overall mortality rate. In this study, we use the tools of Machine Learning (ML) namely, Logistic Regression (LR) to estimate the probability of heart disease through several clinical and physiological variables including age, blood pressure, cholesterol, and heart rate. The main hypothesis of the research is to test the predictive power and clinical interpretability of the Logistic Regression model in order to detect people with the risk of developing heart disease. We used a publicly available heart disease dataset to train and validate the LR model to be able to classify patients as either being at risk or not. The model has an accuracy of 85 which shows that it has a good balance between sensitivity and specificity. According to our results, Logistic Regression can provide a valid, clear, and computationally greedy way of predicting heart disease. Moreover, we ran LR performance against other machine learning models like the Decision Trees, the Random Forests, and the Support Vector Machines. Some of the models had a slight improvement in the accuracy, but LR had a better interpretability- which is critical in clinical decision-making. This study highlights the importance of interpretable ML models in healthcare since they help medical professionals to comprehend risk factors better, assist in diagnostic work, and optimize patient outcomes..
Keywords: Heart disease, machine learning, Logistic Regression, healthcare, predictive modeling, risk assessment, clinical data, machine learning.

ABSTRACT:
Fake review detection is a significant issue in e-commerce where product reviews are significant in consumer decision-making. This paper suggests a hybrid machine learning method that involves textual TF-IDF features and sentiment-based and behavioral Meta features like rating-sentiment gap and review length. The reason behind the use of a Linear Support Vector Machine (SVM) classifier is its performance in the high-dimensional feature spaces. Although text-only models have high in-dataset accuracy (99%), their accuracy decreases dramatically in cross-dataset tests. The hybrid model proposed has 92% accuracy and thus it has good generalization. The findings demonstrate the significance of incorporating linguistic and behavioral aspects to make fake review detection effective.
Keywords: HFake Review Detection, Opinion Spam, TF-IDF, Sentiment Analysis, Hybrid Features, Support Vector Machine.

ABSTRACT:
Breast cancer continues to be among the leading causes of death for women; thus, early detection is critical. We applied three machine-learning classification algorithms, Logistic Regression, Support Vector Machine (SVM), and Random Forest, to classify tumors as either benign or malignant using a common dataset for breast cancer, the Breast Cancer Wisconsin (Diagnostic) dataset. All three algorithms performed remarkably well with accuracy of 96.4% of both SVM and Logistic Regression whereas Random Forest showed 97.3% accurate results, demonstrating how effective classic machine-learning methods can be when used to analyze structured biomedical data. The goal is not only to predict correctly, but to also gain insight into why each algorithm provided its prediction. By using SHAP (SHapley Additive exPlanations), we can analyze the reasons for classification both globally and for particular patients. When examining global SHAP results, features associated with tumor boundary irregularity (e.g., concave points, concavity, worst radius, and worst area), were among the best predictor features across every model. At the patient level, local explanation showed how each feature impacted the overall predicted outcome either positively or negatively. Significantly, the features identified via the model were consistent with what pathologists look for in a malignant tumor and align with actual medical knowledge. The bottom line here is that combining highly effective Machine Learning (ML) algorithms with Explainable Artificial Intelligence (XAI) tools like SHAP provides more accurate predictions and enables physicians to make more informed decisions based on transparent and reliable XAI tools. This type of methodology will create more trustworthy AI systems related to clinical diagnosis as any feature that brings high shift into the prediction will help determine the factors which lead to a particular prediction.
Keywords: Artificial Intelligence, Machine Learning, Explainable AI, Support Vector Machine.

ABSTRACT:
Agricultural management is highly influenced and perfectly organized based on the weather conditions, accurate forecasting helps to increase the crop yield productivity and maintain agricultural activities in a better way. The proposed methodology provides a system that integrates both advanced XGBoost and Long Short-Term Memory (LSTM) models techniques with a multilingual chatbot for forecasting the accurate weather condition and suggests advanced crop monitoring techniques as a recommendation for Precision farming. The Integrated Multilingual chatbot suggests farmers with appropriate recommendations on fertilizer agricultural techniques, irrigation, plants management, and harvesting plans. XGBoost is a trained model particularly used for identification of nonlinear climate patterns for accurate forecasting measures. LSTM monitor and captures temporal dependencies in forecasting weather parameters such as temperature, rainfall, and humidity. This system results show that the integration of both advanced hybrid XGBoost and LSTM techniques achieves higher prediction accuracy and lower weather prediction error than traditional methodology systems. This feature enables improved resource utilization and results in high yeild productivity.
Keywords: Smart Agriculture, Precision Agriculture, Weather Forecasting, XGBoost Model, Hybrid Machine Learning Model, LSTM (Long Short-Term Memory), Deep Learning, Artificial Intelligence in Agriculture, Agricultural Decision Support System, AI Chatbot.

ABSTRACT:
The Static Random Access Memory (SRAM) forms a crucial integrant in embedded systems. Demand for reduced power consumption has made single-ended SRAM bit-cells extremely popular. For enhanced performance of the SRAM bit-cells, the technology node scaling is prevalent to obtain increased transistor density. However, as a consequence the bit-cell undergoes various process-voltage-temperature variations are introduced during the fabrications process. Therefore, the paper studies a single-ended 7 transistor SRAM bit-cell that undergoes 1000-point Monte Carlo simulations at 0.5V VDD under 5% threshold voltage variation at various process corners to measure its read and write stability and to conclude the best and worst performing corner of the bit-cell and evaluate nominal performance values.
Keywords: SRAM cell, process corner, Monte Carlo.

ABSTRACT:
This paper presents an examination of the SCR simulation process through application and comparison of the Ebers-Moll and Gummel-Poon transistor models in a two-transistor configuration. The SCR was simulated in SPICE as two coupled PNP and NPN transistors; the same operating conditions were used to analyze the SCR's switching behavior. The Ebers-Moll model shows very sharp switching because it does not take into account any dynamic effects, whereas the Gummel-Poon model shows smooth transitions with transient oscillation that can be attributed to the charge storage effects and junction voltage effects considered by this model. Both models showed a similar time for switching when the voltage collapsed however, the Gummel-Poon model better represents the real-world SCR dynamics.
Keywords: BJT, Silicon Controlled Rectifier (SCR), Two-Transistor Model, Ebers-Moll Model, Gummel-Poon Model, SPICE Simulation.

ABSTRACT:
Biometric recognition has emerged as a reliable method for enhancing security in modern digital and physical systems. Authentication using fingerprints has emerged as a reliable method for enhancing security in modern digital and physical systems. The Arduino Uno microcontroller, Arduino IDE, and GT521-F32 fingerprint sensor module were used to create fingerprint recognition system presented in this paper. The design architecture, system workflow, and hardware as well as software integration were all reviewed in this paper. Additionally, reviewed were the system's accuracy, cost efficiency, and real-time performance. The review also discusses major challenges, such as insufficient storage capacity, sensitivity to the environment and security vulnerabilities. Additionally, the most recent developments in the field such as IoT-based authentication and intelligent biometric systems—are discussed to outline the advancements that will be made in the future. The results imply that fingerprint recognition with Arduino boards are a viable and economical alternative to bigger systems and can be enhanced in scalability and resilience. This project is aimed at offering an academic study as well as a procedural guide to the electrical characteristics, the specification of UART, the matching and enrolled individuals based algorithms on the module, and how to implement the module practically using Arduino.
Keywords: Fingerprint Recognition, Biometric Authentication, Arduino Uno, GT521-F32 Sensor, Embedded Systems and IoT-Based Authentication.

ABSTRACT:
The rapid emergence of Internet of Things (IoT) applications has made it even more imperative to schedule tasks efficiently. The traditional algorithms, e.g., First-Come-First-Serve (FCFS) and Earliest Deadline First (EDF) algorithms are concerned mainly with the time ordering of tasks, and do not take into account the energy constraints and varying load conditions in hierarchical systems. In this paper, we present an energy-efficient and priority-based task scheduling framework for hierarchical IoT networks that include edge, fog and cloud layers. The aim of the proposed scheme is to balance queue load across hierarchical IoT layers while considering edge node energy consumption and task priority to ensure efficient and reliable task scheduling. To model task arrivals and service processes under varying load conditions, M/M/1 and M/M/c queueing models are employed for performance analysis and adaptive scheduling decisions. Simulation results demonstrate that the proposed approach outperforms FCFS and EDF in terms of reduced response time, energy consumption, and enhanced overall system efficiency. These findings highlight the effectiveness of integrating hierarchical task management, queueing analysis, and adaptive priority scheduling in modern IoT systems.
Keywords: Internet of Things, Energy-aware scheduling, Priority-based task scheduling, Load balancing, Service Level Agreement.

ABSTRACT: Skin cancer, being a worldwide health problem, creates major physical, mental and emotional stress. Traditional physical finding is majorly time-taking and subjective, depending on much upon dermatologists' knowledge. In our work, an extensive skin cancer disease analysis is proposed for ease in diagnosis by using deep learning models. The HAM10000 dataset is the standard for estimating skin cancer disease by AI detection approaches. It consists of more than 10,015 dermoscopic images labeled with seven different diagnostic categories. We have identified the skin cancer ResNet18 to the HAM10000 dataset. The procedures in detail, experimental setup, results, and conclusions include reviewing findings and potential consequences on the clinical procedure.
Keywords: skin cancer, CNN, ResNet, HAM100000 Dataset.

ABSTRACT: Neuromorphic computing moves away from the Von Neumann architecture. Instead of developing computers with traditional architecture, neuromorphic computing imitates how the brain could potentially operate by incorporating the construction of neurons and synapses. This can be attributed to the impasses of Moore's Law, the "memory wall," and the demands of traditional computer architecture. This review paper delves into the evolution of computers from their conception to current artificial intelligences and the limits of traditional computers, particularly the "Von Neumann bottleneck." The review demonstrated the differences between basic biology and artificial constructs of neurons. This was used to show how neuromorphic computing and the incorporation of Spiking Neural Networks (SNNs) operate with long-term energy efficiency and large-scale parallelism. Architectures mimicking the brain with synaptic plasticity were termed the "Silicon Brain." This review also encompassed contemporary neuromorphic computing architectures such as Loihi and Loihi 2, TrueNorth, and SpiNNaker2, in addition to memristor technology. This review also elaborated on the various applications of neuromorphic computing in autonomous vehicles, cyber security, Edge AI, and neuromorphic computing's application in healthcare, while also addressing the barriers of use, such as the difficulties of programming and maintaining accuracy in computations. It was concluded that neuromorphic computing holds the greatest potential for the future of energy-efficient computing.
Keywords: Neuromorphic Engineering, Spiking Neural Networks (SNN), Von Neumann Bottleneck, CMOS, Silicon Brain, Edge AI, Event-driven processing, Loihi, Loihi 2, TrueNorth, SpiNNaker2, Memristor, Spike-Timing-Dependent Plasticity (STDP), In-Memory Computing.

ABSTRACT: Degraded fingerprints caused by dry skin, dirt, or weak sensors remain a significant problem in biometric authentication. To improve the efficiency of neuromorphic recognition, this paper proposes a two-stage framework that integrates fingerprint enhancement with neuromorphic recognition. A Dense Attention U-Net (DenseAttUNet), guided by ORRAM, is used to restore damaged ridge patterns. The restored output is subsequently injected into an EfficientNet-B2 backbone, where ERRAM bridges spatial characteristics with a Spiking Neural Network (SNN) for energy-efficient classification. The entire system was trained and tested on 4,695 degraded fingerprints. The enhancement module achieved a PSNR of 18.04 dB and SSIM of 0.7380. Recognition achieved 85.12% accuracy and an F1 score of 85.74%. These findings verify that combining brain-inspired computing with intelligent image restoration provides a robust, scalable fingerprint recognition system under real-world degradation.
Keywords: Fingerprint Enhancement, Spiking Neural Network (SNN), Dense Attention U-Net, EfficientNet-B2, Biometric Recognition, Neuromorphic Computing.

ABSTRACT: With the emergence of renewable energy sources and the decrease in system inertia, islanded microgrids face stability challenges. Grid-following inverters cannot control voltage and frequency independently in such circumstances. In this paper, a droop control based Virtual Synchronous Generators (VSGs) is proposed to simulate synchronous generator’s behavior by offering virtual inertia and damping. For the regulation of a three-phase voltage source inverter with an output filter, a dq-frame voltage-current dual-loop controller is used in conjunction with P-f and Q-V droop loops. MATLAB/Simulink was used to assess the model's performance under islanded operation. The system demonstrates fast dynamic response with a frequency overshoot of less than 0.04% and rapid convergence to the nominal value of 50 Hz, ensuring stable and well-damped operation. A good damping in the power output characteristics, voltage regulation accuracy, and sinusoidal waveform balance are exhibited from the simulation outcomes. The suggested technique guarantees stable voltages and frequencies in applications with low inertia microgrids.
Keywords: Droop control, dq control, Grid-forming inverter, Islanded microgrid, Virtual Synchronous Generator.

ABSTRACT: Breast cancer is a major health concern in the world, and requiring early and accurate diagnosis in order to be treated. Due to the improved artificial intelligence, deep learning approaches have been used. Have demonstrated much potential in medical image analysis through giving automated and reliable diagnostic support. However, single-model processes can easily struggle to locate the complex. Patterns that are in medical data. This study proposes a hybrid deep learning model. for breast cancer detection through feature-level fusion of convolutional neural networks. The technique employs DenseNet and EfficientNet along with it to extract various features on input images. This allows the model to learn deep and efficiently scaled. Representations. We also created a model, which compares ResNet. and Inception architectures to find out how well various hybrid combinations work.
In order to make the system more stable and generalize, it takes advantage of. data augmentation and nowmalization preprocessing techniques. It is designed to classify pictures into three categories, normal, malignant, and benign. The findings indicate that feature-level fusion. is effective in enhancing model performance and might be used to assist with. early diagnosis. This work stresses how important hybrid architectures are for medical imaging and preconditions the further development of the smart breast cancer detecting systems.
Keywords: Breast Cancer, Deep Learning, Convolutional Neural Networks (CNN), Feature-Level Fusion, DenseNet, EfficientNet, Medical Image Analysis.

ABSTRACT: The digital platforms have brought a revolution to the way information is stored and shared. But the websites have revealed the weaknesses which it possesses in the old traditional, centralised kind of storage. All the recent news about privacy invasion, manipulations of data, server failures, and a single source of control over the key data to a user has raised the possibility of more demands towards other alternatives that resulted in the emergence of Blockchain and the Inter Planetary File System (IPFS). These technologies are decentralising trust and storage, which is to allocate the trust and storage in multiple nodes. Blockchain is intended to ensure a standardized path of the impartial history of transactions that cannot be changed; whereas the IPFS offers a more prosperous procedure of allocating files by offering the user the capability to relate the files through content addressing. Consequently, this paper is the proposal of the synergistic model of Blockchain and IPFS deploying Smart Contracts to be able to verify the level of user access, control the Metadata affixed to files, and also make interactions super-resistant to tampering. The proposed Hybrid framework is better in terms of integrating the data integrity, less dependency on centralized servers or more resilience in relation to unauthorised entry to the data store of vital information or system malfunction. Whereas latency when using decentralised peer-to-peer networks may be a bit higher than when making use of the conventional centralised servers, the decentralised networks are more transparent, verifiable and secure when storing data over the long run. The Hybrid Blockchain-IPFS model proves that new emerging modern applications that need reliable censorship-resistant and secure sharing of files can use the decentralised storage.
Keywords: Blockchain, InterPlanetary File System (IPFS), Decentralized Storage, Censorship-Resistant Storage, Decentralized Applications (DApps).

ABSTRACT: Early level fall detection is crucial for health care surveillance, geriatric medicine, and public safety. This paper introduced real-time fall detection algorithm that monitors joints of the body with YOLOv8n pose estimation and assess their score with our novel biomechanics approach. The program offers four video source selections: Webcam, IP/USB camera, Stored video files or RTSP/HTTP stream. To detect a fall, five things are taken care simultaneously: how the box shape is altered, how much the torso tilts, where the head sits versus that of the hips, how close knees are to shoulders, and how high ankles rise. You only get alerts if the additional evidence leads to S≥4, so false alarms drop a lot. Run on your own local machine as a flask app, streams live MJPEG, overlays skeletons, lights up alerts instantly and record starts for every session. This system achieves 94.3 % sensitivity and 96.9 % specificity on the UR Fall Detection and Multiple Cameras Fall datasets, with an overall accuracy of 95.4 % + a false alarm rate of only 3.1 %, outperforming single-energy baselines. The YOLOv8n-pose backbone inference ~28–32 FPS on a standard home CPU, thus it can be deployed to edge without a GPU.
Keywords: YOLOv8, human fall detection, health care, UR Fall Detection.

ABSTRACT: The paradigm of contrastive learning has become an effective representation learning method in natural language processing, using semantic similarity of the augmented data pairs to enhance model robustness. This paper presents empirical research on the use of contrastive learning- namely the NT-Xent (Normalized) Cross Entropy) loss-scaled by temperature-to a large scale multidomestic twitter sentiment dataset of 74,682. Tweets in 32 brand and game categories and four sentiments. One of the distinguishing characteristics of this dataset is. Its 6x augmentation, where a single original tweet is paraphrased six times with the same paraphrase exists. Identifier, which has been shown to give rich positive pairs on which to train contrastively. Our hypothesis is that we will suggest a dual-objective BERTweet. Based architecture which consists of cross-entropy classification and NT-Xent contrastive regularization (=0.5, tau=0.07), and make a systematic comparison with a powerful non-augmented baseline. Our, contrary to expectation, is as follows: contrastive model has a Macro- F1 of 0.6026 in terms of Test as against the baseline of 0.6496- a decrease of 4.70 points. We prove this degradation not to be uniform, showing, by detailed per-class analysis, that it is the Irrelevant class that is harmed. the most significant decrease (F1: 0.49 to 0.36), whereas classes that carry sentiments are not very volatile. And we ascribe this to three. compounding factors: (1) augmentation faithfulness violations in the dataset which add label incompatible variants, (2) lack of batch diversity in optimizing NT-Xent on CPU hardware that has a bottleneck on in-batch negative richness, and (3) the irrelevance of the Irrelevant class as a contrastive target. The first is our findings. Formal investigation of contrastive learning with augmentation support on this popular benchmark, and our analysis of failure. Offers practical principles on how practitioners can use contrastive goals when using noisy, multi-class Twitter corpora. We also are able to show that contrastive objective is selectively detrimental to underrepresented and semantically ambiguous. Classes, which have a direct consequence on fairness-conscious NLP systems. Index Terms used- Sentiment analysis, contrastive learning, Twitter NLP, data augmentation, BERTweet, NT-Xent, multi-domain classification, representation learning.
Keywords: Sentiment analysis, contrastive learning, Twitter NLP, data augmentation, BERTweet, NT-Xent, multi-domain classification, representation learning.

ABSTRACT: In environments with high density such as religious events, transport hubs and large gatherings, stampede events can cause many deaths. If stampede events could be predicted and analyzed early on, there could be less fatalities and better capability to respond to emergencies. This paper describes a proposal for using machine learning methods to detect the risk of a stampede using CSV structured data (from historical stampede incidents) for analysis. In particular, this study used the Human Stampedes dataset which contains information on stampedes in world history dating from 1800– 2021. The study applied three different supervised machine learning models including Random Forest, XGBoost and Logistic Regression to classify the risk (or hazard) level of a crowd as part of the stampede detection process. In order to improve the interpretability and trust in the model results, the study also used Explainable Artificial Intelligence (XAI) techniques (SHAP, LIME) to develop an explanation for each model’s prediction. The combination of predictive accuracy and transparency improves the success of the overall proposed model, and thus allows it to be used in smart surveillance systems.
Keywords: Stampede detection, crowd analysis, machine learning, explainable AI, SHAP, LIME, CSV dataset, public safety.