Your search

Completely Automated Public Turing tests to tell Computers and Humans Apart (CAPTCHAs) are widely used mechanisms for protecting websites against automated bot attacks. In this paper, we investigate the trade-off between usability and security in CAPTCHA design using a Stackelberg game-theoretic framework. Based on empirical evaluation data on CAPTCHA solving time and accuracy, we incorporate expected solving time into the defender's utility function to jointly capture usability and security. We propose two models: a discrete-strategy game, in which different CAPTCHA types and difficulty levels form the strategy space, and a continuous-strategy game, in which CAPTCHA hardness is represented as a continuous decision variable. For the discrete model, we derive a stepwise solution characterizing transitions in the defender's optimal strategy. For the continuous model, we analyze a threshold-based attacker response and obtain equilibrium solutions under deterrence and accommodation regimes. The proposed framework is adaptable to other CAPTCHA evaluation datasets.

This study evaluates the viability of deep learningbased super-resolution (SR) for enhancing real-time video streaming. We compare a traditional HEVC-compressed streaming pipeline against an AI-assisted framework that transmits lowresolution video and reconstructs high-resolution output at the client using the Efficient Sub-Pixel Convolutional Neural Network (ESPCN) algorithm. The evaluation spans multiple resolutions (1080p to 4K), motion types, and processing methodologies (subprocess vs. frame-by-frame). Experimental results indicate that while AI-assisted streaming significantly reduces transmission bandwidth, it introduces substantial computational overhead, with completion times increasing by over 300% compared to traditional methods. Furthermore, quality assessments using PSNR and VMAF metrics reveal that most AI-assisted streams failed to meet the target quality range (30–50 dB PSNR), with high-resolution 4K streams experiencing up to 24.58% frame loss during real-time inference. These findings highlight critical performance barriers in utilizing current SR algorithms for standard consumer-grade real-time video delivery.

Capacitive swipe gestures have recently emerged as a promising behavioral biometric modality for user authentication and identification on smartphones. While prior studies have demonstrated their effectiveness using classical machine learning techniques, the potential of deep learning in this domain remains largely unexplored, particularly under limited training data conditions. In this paper, we investigate the use of deep learning for capacitive swipe–based user identification and address the challenge of data scarcity through an image fusion strategy. Experiments on a dataset of 30 users show that the proposed "deep learning plus image fusion" approach significantly improves performance over a standalone deep learning baseline and achieves results comparable to state-of-the-art classical machine learning methods. These findings highlight the importance of data size and demonstrate the potential of image fusion for enhancing deep learning–based user identification.

Background/Objectives: Obstructivesleep apnea (OSA) is one of the most widespread forms of sleep disease, affecting over 936 million adults globally. The health consequences of obstructive sleep apnea (OSA) are well documented; however, it remains largely underdiagnosed because the current gold-standard diagnostic method, polysomnography (PSG), is often costly, time-consuming, and unavailable in many healthcare settings. To address these challenges, this study presents a novel explainable deep learning (DL) framework for automated multi-class OSA severity classification using single-lead electrocardiogram (ECG) signals. Methods: The proposed framework integrates a hybrid CNN–BiLSTM architecture with explainable artificial intelligence (XAI) techniques to generate clinically meaningful predictions and explanations across four OSA severity classes: Normal, Mild, Moderate, and Severe. The framework was evaluated using the publicly available PhysioNet Apnea-ECG dataset (70 recordings) together with an institutional ECG dataset (150 recordings), resulting in a combined cohort of 220 recordings. Results: The proposed framework achieved an overall classification accuracy of 94.7%, with sensitivity and specificity values of 92.3% and 96.1%, respectively. Furthermore, the proposed model consistently outperformed conventional machine learning algorithms, including Support Vector Machine (SVM), Random Forest, and XGBoost, by 5.5%, 4.2%, and 2.9%, respectively. To enhance transparency and clinical trust, SHAP (SHapley Additive exPlanations) was employed to identify the most influential physiological predictors driving model decisions. Heart rate variability features, particularly RMSSD and pNN50, emerged as the strongest indicators of OSA severity. Moreover, computational efficiency analysis revealed that the model required only 0.23 s to process a 60 s ECG epoch on a standard computing platform, supporting its suitability for real-time deployment. Conclusions: The findings demonstrate that explainable deep learning applied to ECG signals can provide accurate, interpretable, and computationally efficient assessment of OSA severity. The proposed framework may support OSA screening, clinical triage, and early intervention, particularly in resource-constrained healthcare environments.

This research presents a comparative performance analysis of three prominent hypervisor architectures—KVM, Xen, and VMware Workstation—to evaluate their scalability and resource efficiency under increasing virtual machine (VM) densities. Utilizing a standardized benchmarking framework, this study quantifies performance regressing across CPU throughput (GFLOPS), memory bandwidth, disk I/O, and initialization latency. The experiment results reveal that architectural design significantly dictates scalability limits: KVM’s kernel-integrated Type-1 model achieved the highest efficiency, maintaining a near-native CPU baseline and superior I/O resilience at peak density. In contrast, Xen’s microkernel architecture demonstrated stable but lower throughput due to Domain 0 (Dom0) management overhead, while VMware’s Type-2 hosted model exhibited the most acute performance degradation. These results demonstrate that while hosted hypervisors are suitable for low-density personal use, kernel-level integration is a mechanical necessity for maintaining performance in high-concurrency server environments. This work provides a reproducible methodology and critical data for system architects and researchers seeking to optimize resource orchestration in distributed and parallel systems.

Arabic extractive text summarization remains challenging due to rich morphology and limited lexical consistency, which weaken traditional surface-based sentence ranking methods. Existing approaches typically rely either on graph centrality to capture document structure or on transformer-based embeddings to estimate semantic relevance, yet the relationship between these two ranking signals has not been systematically examined. This paper investigates the complementary roles of semantic similarity and structural centrality in Arabic summarization. A controlled hybrid ranking formulation is implemented that combines contextual sentence embeddings with PageRank-based graph scoring and evaluate it using multi-reference ROUGE on the Extended Arabic Summaries Corpus (EASC). The study analyzes how each component influences sentence salience, coverage, and redundancy. Experimental results show that semantic embeddings improve the identification of informative sentences, while graph centrality enhances coverage and reduces repetition; their integration consistently yields stronger summaries, achieving ROUGE-1 = 0.605, ROUGE-2 = 0.497, and ROUGE-L = 0.514. These findings provide empirical evidence that semantic and structural centrality capture complementary aspects of importance in morphologically rich languages, offering guidance for designing more robust extractive summarization systems.

Background/Objectives: Obstructive sleep apnea (OSA) is a common and serious sleep-related disorder that causes repeated interruptions in breathing during sleep. Traditional diagnostic methods, such as polysomnography, are accurate but costly, time-consuming, and unsuitable for large-scale screening. This study proposes and evaluates a lightweight diagnostic framework based on an Extreme Learning Machine (ELM) optimized by a set of basic and advanced metaheuristic optimizers. The model aims to evaluate whether metaheuristic optimization can improve ELM-based classification performance using structured demographic, clinical, and sleep-related predictors. Methods: Two real datasets were employed to train and evaluate the proposed framework: (i) a clinical OSA dataset with 274 subjects and 31 demographic/anthropometric and sleep-related predictors, and (ii) a public strongly imbalanced Sleep-Disordered Breathing (SDB) dataset with 500 subjects and 10 structured predictors. Metaheuristic algorithms are used to optimize ELM weights and biases, addressing the instability of random initialization and improving model generalization. The optimized models are evaluated against eight baseline classifiers, including logistic regression (LR), k-nearest neighbors (KNN), decision tree (DT), random forest (RF), support vector machine (SVM), multilayer perceptron (MLP), XGBoost (XGB), and a standard ELM classifier. Results: Results show that metaheuristic optimization moderately improves ELM on the OSA dataset, increasing ROC-AUC from 0.6527 to about 0.73 and accuracy from 0.6573 to about 0.69–0.70, while on the highly imbalanced SDB dataset, it yields modest ROC-AUC gains (from 0.5132 to about 0.544–0.548) with small decreases in accuracy and F1-score. We additionally assess class-imbalance handling on the SDB dataset and analyze feature importance with permutation importance and SHAP, which shows the models rely heavily on diagnosis-derived predictors. Conclusions: The proposed framework provides a lightweight ELM-based decision-support approach with low inference cost after offline optimization. The results suggest potential value for screening-oriented OSA/SDB classification, but further validation with larger cohorts and a screening-only feature set is needed before clinical implementation.

Cardiovascular diseases (CVDs) are still one of the foremost causes of death globally. An urgent need for an accurate premature diagnosis is essential for efficient treatment and avoidance. The goal of this research is to introduce our initial idea for developing a hybrid intelligent diagnosis model that integrates the Extreme Learning Machine (ELM) with Particle Swarm Optimization (PSO) to enhance diagnosis accuracy for heart disease. To address the problems of random initialization, the PSO algorithm is used to optimize the ELM’s input biases and weights. Experiments using the Cleveland Heart Disease dataset evaluated the performance of traditional ELM and the proposed PSO-ELM model. Results indicate that integrating PSO for parameter tuning yields significant improvements across evaluation metrics. The traditional ELM recorded an accuracy of 0.8595 on the training set and 0.8525 on the test set, while the PSO-ELM model achieved up to 0.9174 in training and 0.8852 in testing. Enhancements were also noted in precision, recall, and F1-score, with the F1-score increasing from 0.8768 for traditional ELM to over 0.92 for several PSO configurations.

Gold serves as a safe-haven asset, an inflation hedge, and a store of value, among other economic purposes. History has a wealth of information about its connections to important macroeconomic and financial variables. We compare three supervised methods for one-day GLD (SPDR Gold Shares) prediction using cross-Asset drivers: SPX (S&P 500 Index), USO (United States Oil Fund), SLV (iShares Silver Trust), and EUR/USD (Euro-U.S. Dollar exchange rate). To fill this gap: (i) a linear regression baseline, (ii) a feedforward neural network (FFNN) trained by a Genetic Algorithm (GA), and (iii) an FFNN optimized through Particle Swarm Optimization (PSO). We use standard scaling, a held-out train/test split, and R2, Variance Accounted For (VAF), MSE, RMSE, MAE, and GA/PSO convergence curves to check how the optimizer works with daily data from 2008 to 2018. Empirically, PSO-FFNN does better than both the linear baseline and GA-FFNN. It has the best generalization (for example, Test MSE 38.99, R 0.929). This implies that (1) GLD possesses exploitable nonlinear structure concerning these predictors, and (2) PSO traverses the FFNN search space more effectively than GA in our context. The findings validate the application of evolutionary training for reliable and accurate gold price forecasts, with implications for risk management and strategic asset allocation. © 2025 IEEE.

Water flow prediction and planning significantly help decision-makers determine the most suitable irrigation strategy and crop type and help avoid risks from flooding, among other benefits. Conventional statistical and physical models are often challenged by the highly dynamic and nonlinear nature of hydrological processes. Recent advances in machine learning (ML), including artificial neural networks (ANNs), provide powerful tools for modeling these complex relationships. However, the performance of these models depends on the optimal parameter tuning. By combining ANN with Crow Search Optimization, we aim to improve prediction accuracy and robustness while providing a clever, adaptable, and reliable solution to real-world water flow forecasting problems. The Crow Search Algorithm (CSA) is one of the most recent metaheuristic algorithms used as a training algorithm for neural network models to achieve higher performance. This research provides an evolutionary-based model to predict the flow of the Black River, a well-known river in the USA. The adopted ANN model was used to train and predict daily flows at the initial Black Water River station (No. 02047500) near Dendron, Virginia. Among the well-known metaheuristic algorithms employed in this study for comparison are the Salp Swarm Algorithm (SSA), Particle Swarm Optimization (PSO), and the Dandelion Optimizer (DO). Based on comparative research, the CSA algorithm outperforms other training algorithms in predicting river flow, achieving an average fitness value of 0.0048926, which is 41% better than SSA, 81% better than PSO, and 49% better than DO. Furthermore, CSA has achieved a superior convergence curve, and high variance accounts for VAFs of up to 99.06% on the training data and 98.45% on the test data. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Abstract Description The identification and prioritization of cancer-specific neoepitopes from next-generation sequencing data for personalized immunotherapies such as cancer vaccines remains challenging and requires the use of complex bioinformatics approaches. Here, we present GeNeo2, an updated version with enhanced features of the GeNeo toolbox for predicting neoepitopes from matched tumor/normal exome sequencing data coupled with tumor RNA-Seq data (Al Seesi et al., 2023). Unlike GeNeo, which identifies neoepitopes generated by single nucleotide variants, GeNeo2 also predicts neoepitopes generated by somatic indels. A distinguishing feature in GeNeo2 is that it integrates tools for analyzing mass spectrometry immunepeptidomic data, which can reveal neoantigens derived from both canonical and noncanonical sources. Finally, GeNeo2 integrates novel machine-learning approaches to improve the accuracy of somatic variant calling and peptide identification from mass spectrometry data. GeNeo2 tools can be accessed via web-based interfaces deployed on a Galaxy portal accessible at https://neo.engr.uconn.edu/. A virtual machine image for running GeNeo2 locally is also available to academic users upon request. Topic Categories Computational and Systems Immunology (COMP)

Insertion or deletion of one or two base pairs within a coding region causes a frameshift, which has the potential to generate neoepitopes (InDel-generated neoepitopes) that lack a self-counterpart and are entirely novel. Despite the obvious appeal of InDel-generated neoepitopes, and the demonstration of such candidate neoepitopes that can elicit a CD8 T-cell response, no InDel-generated neoepitopes that actually control tumors in vivo have been reported thus far. Here, in a mouse colon carcinoma line, we identify 11 InDels, only one of which generates a neoepitope that elicits tumor control in vivo in models of prophylaxis as well as therapy. Although this neoepitope has no self-counterpart, it has a low affinity (IC50 33,937.60 nM) for its MHC I allele. Despite its low affinity for MHC I, this neoepitope elicits antitumor activity in vivo through CD8 T cells. Furthermore, CD8 T cells elicited by this InDel-generated neoepitope, like the neoepitopes created by point mutations, show notably less exhaustion than classical immunogenic epitopes. Ironically, this InDel-generated neoepitope follows the same rules as noted for most of the tumor control-mediating neoepitopes generated by point mutations that have a poor affinity for MHC I alleles.

Proportional-Integral-Derivative (PID) controllers are prominent due to their superior functionality and ease of use. However, optimizing their parameters presents a significant challenge. Adjusting parameters must be done carefully and cautiously because improper calibration can compromise the system’s stability. Although classic tuning techniques, such as the Ziegler-Nichols (ZN), are frequently employed, their efficiency is restricted due to the intricate and ever- changing nature of the systems, often leading to parameter settings that could be more optimal. Therefore, the need for a more accurate parameter-tuning technique is urgent. Various optimization strategies are used to fine-tune parameters with more precision. These methods include Gray Wolf Optimization (GWO), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO). These methods are applied to fine-tune the PID parameters for a Direct Current (DC) motor to achieve optimal performance, and a comparative analysis of the results is conducted. Various fitness functions encompass performance metrics such as rise time, overshoot, peak time, settling time, and mean square error (MSE). These metrics are incorporated into the corresponding optimization approaches to quantitatively assess the controller’s performance. Various test cases have been utilized and the GA outperforms other algorithms ranging from 17% to 28% where rise time, settling time, and MSE are significant in the fitness function.

The increasing volume of suspicious emails, commonly known as spam, has created a critical need for more reliable and robust anti-spam filters. These suspicious emails can be dangerous and can lead to the loss of personal information, underscoring the necessity for an effective spam filtering system. The application of machine learning methods has enhanced system security and improved the detection of suspicious messages. This research evaluates the effectiveness of seven machine learning algorithms for classifying suspicious email messages: random forest, support vector machine, artificial neural network, decision tree, gradient boosting classifier, and k-nearest neighbor. The primary focus of this evaluation is the accuracy achieved by each algorithm in identifying spam emails. Our analysis revealed that the random forest algorithm outperformed the other evaluated algorithms in terms of accuracy for spam email classification, achieving a remarkable 95%. The accuracy percentages of the various methods ranged from 88% to 93%. Copyright 2025. The Korean Institute of Information Scientists and Engineers.