Journal of Computing Theories and Applications

The Llama–ARCS Adaptive Learning framework: AI–VR Integration System for Real-Time Motivational Feedback in Higher Education

Abraham Eseoghene Evwiekpaefe, Darius Tienhus Chinyio, Loreta Katok Tohomdet — Mon, 23 Feb 2026 00:00:00 +0000

This study developed and evaluated an AI-integrated Virtual Reality (VR) system designed to enhance personalized learning in higher education. While VR improves engagement, existing systems often lack adaptivity or experience high latency during AI interactions. To address these limitations, this research introduces a novel integration of a cache-optimized Llama 2 Large Language Model (LLM) that delivers real-time, motivationally grounded feedback. The system was implemented using Unity 3D and validated with 50 undergraduate students. Technical validation showed that the cache layer reduced interaction latency from 17.7 ms to 14.2 ms and maintained zero system crashes throughout the pilot. Learner motivation was assessed using Keller’s ARCS model, yielding mean scores ranging from 4.08 to 4.69 across all dimensions. Independent t-tests (p > 0.05) and negligible effect sizes (Cohen’s d < 0.2) revealed no significant difference between technical (ICT) and non-technical (Physics) students. These findings confirm that the proposed system effectively bridges technological and motivational gaps, providing a robust model for adaptive, immersive education.

Integrating Fully Homomorphic Encryption and Zero-Knowledge Proofs for Efficient Verifiable Computation

UmmeAmmara Qureshi, Bhumika Doshi, Aditya More, Kashyap Joshi, Kapil Kumar — Mon, 23 Feb 2026 00:00:00 +0000

Fully Homomorphic Encryption (FHE) enables computation on encrypted data with end-to-end confidentiality; however, its practical adoption remains limited by substantial computational costs, including long encryption and decryption times, high memory consumption, and operational latency. Zero-Knowledge Proofs (ZKPs) complement FHE by enabling correctness verification without revealing sensitive information, although they do not support encrypted computation independently. This study integrates both techniques to enable encrypted computation with verifiably consistent results. A prototype system is implemented in Python using Microsoft SEAL for homomorphic encryption and PySNARK for Zero-Knowledge Proof verification. Experiments are conducted on standard consumer-grade hardware (Intel i5, 8 GB RAM, Ubuntu 22.04) using datasets ranging from 100 MB to 1 GB. The evaluation focuses on encryption and decryption time, homomorphic computation latency, memory usage, and proof generation overhead. Experimental results show that integrating ZKPs introduces a moderate and stable runtime overhead of approximately 15–20%, as analyzed in Section 4, while enabling verification without plaintext disclosure. Ciphertext expansion remains a notable limitation, with observed growth of approximately 30–40× relative to plaintext size, consistent with prior FHE implementations. Despite these overheads, the system demonstrates feasible scalability for datasets up to 1 GB on mid-level hardware. Overall, the results indicate that the integrated FHE+ZKP approach provides a practical balance between confidentiality, verifiability, and performance, supporting its applicability to privacy-preserving scenarios such as secure cloud computation, encrypted data analytics, and confidential data processing under realistic resource constraints.

Hybrid Real-time Framework for Detecting Adaptive Prompt Injection Attacks in Large Language Models

Chandra Prakash, Mary Lind, Elyson De La Cruz — Mon, 23 Feb 2026 00:00:00 +0000

Prompt injection has emerged as a critical security threat for Large Language Models (LLMs), exploiting their inability to separate instructions from data within application contexts reliably. This paper provides a structured review of current attack vectors, including direct and indirect prompt injection, and highlights the limitations of existing defenses, with particular attention to the fragility of Known-Answer Detection (KAD) against adaptive attacks such as DataFlip. To address these gaps, we propose a novel, hybrid, multi-layered detection framework that operates in real-time. The architecture integrates heuristic pre-filtering for rapid elimination of obvious threats, semantic analysis using fine-tuned transformer embeddings for detecting obfuscated prompts, and behavioral pattern recognition to capture subtle manipulations that evade earlier layers. Our hybrid model achieved an accuracy of 0.974, precision of 1.000, recall of 0.950, and an F1 score of 0.974, indicating strong and balanced detection performance. Unlike prior siloed defenses, the framework proposes coverage across input, semantic, and behavioral dimensions. This layered approach offers a resilient and practical defense, advancing the state of security for LLM-integrated applications.

Android Malware Detection Using Machine Learning with SMOTE-Tomek Data Balancing

Mon, 23 Feb 2026 00:00:00 +0000

This study presents a comprehensive comparative analysis of four traditional machine learning algorithms Decision Tree, Random Forest, K-Nearest Neighbors, and Support Vector Machine for Android malware detection using the preprocessed TUANDROMD dataset comprising 4,465 instances and 241 features representing both static and dynamic application characteristics. Motivated by the limitations of conventional signature-based and hybrid detection methods, especially in managing imbalanced datasets and detecting emerging malware variants, the study employed SMOTE to ensure balanced training data and fair model evaluation. The dataset was divided into 80% training and 20% testing subsets, and models were assessed using key performance metrics including accuracy, precision, recall, F1-score, and ROC AUC. The findings revealed that the proposed Random Forest model outperformed the other classifiers, achieving an accuracy of 0.993, precision of 0.992, recall of 1.000, F1-score of 0.996, and a near-perfect ROC AUC of 0.9998 surpassing state-of-the-art approaches. These results affirm the superior predictive capability, consistency, and robustness of the Random Forest algorithm in Android malware detection. The study concludes that base models, when integrated with class-balancing techniques, provide reliable and efficient malware detection across imbalanced datasets. For future research, the study recommends exploring advanced hybrid or ensemble frameworks that integrate Random Forest with deep learning architectures or other meta-heuristic optimization techniques to further enhance detection accuracy, adaptability, and resilience against rapidly evolving Android malware threats.

Multimodal Deep Learning for Pneumonia Detection Using Wearable Sensors: Toward an Edge-Cloud Framework

Emmanuel Onwako Ibam, Johnson Bisi Oluwagbemi — Mon, 23 Feb 2026 00:00:00 +0000

Pneumonia remains a leading cause of morbidity and mortality worldwide, particularly in resource-limited settings and among elderly populations, where timely diagnosis and continuous monitoring are often constrained by limited clinical infrastructure. This study presents an edge–cloud–integrated framework for early pneumonia risk monitoring, leveraging multimodal wearable sensors and deep learning to support continuous short-duration monitoring. The proposed system is designed to operate in near real time under simulated deployment conditions, continuously acquiring and analyzing physiological signals (respiratory rate, heart rate, SpO₂, and body temperature) alongside event-driven acoustic biomarkers (cough sounds) within a distributed architecture. A lightweight edge module performs local signal preprocessing and anomaly triage, selectively transmitting salient information to a cloud-based multimodal deep learning model for refined risk estimation and interpretability analysis. The framework was evaluated using a multi-source dataset comprising public repositories (MIMIC-III and Coswara) and a clinically supervised wearable study conducted in two Nigerian hospitals, resulting in 718 hours of quality-controlled multimodal monitoring data. In a pooled multi-source evaluation, the system achieved an AUC of 0.95, while in a clinically realistic local-only evaluation, the AUC was 0.86, reflecting a consistent but preliminary diagnostic signal. These results highlight the importance of local data adaptation for real-world applicability and suggest that multimodal AI can provide meaningful early risk indicators under resource constraints. Beyond predictive performance, this work demonstrates the feasibility of integrating multimodal learning, edge–cloud computation, and explainable analytics into a deployment-aware, privacy-preserving monitoring framework for low-resource healthcare environments.

A Lightweight Maize Leaf Disease Recognition Using PCA-Compressed MobileNetV2 Features and RBF-SVM

Mustapha Abubakar, Yusuf Ibrahim, Ore-Ofe Ajayi, Sani Saleh Saminu — Mon, 23 Feb 2026 00:00:00 +0000

The integration of Artificial Intelligence (AI) into precision agriculture has significantly improved plant disease recognition; however, many existing deep learning models remain computationally expensive and feature-redundant, limiting their deployment on low-power and edge devices. To address these limitations, this study proposes a lightweight framework for maize leaf disease recognition based on serial deep feature extraction, dimensionality reduction, and machine-learning–based classification. A pre-trained MobileNetV2 network is employed as a fixed feature extractor to obtain discriminative visual representations, while Principal Component Analysis (PCA) is applied to reduce feature dimensionality by approximately 76%, retaining 95% of the original variance and improving computational efficiency. The compressed features are subsequently classified using a Radial Basis Function Support Vector Machine (RBF-SVM), optimized via grid search and cross-validation. Experiments conducted on a four-class maize leaf disease dataset (Northern Leaf Blight, Common Rust, Gray Leaf Spot, and Healthy), with class imbalance handled during training, demonstrate that the proposed MobileNetV2–PCA–SVM pipeline achieves 97.58% accuracy, 96.60% precision, 96.59% recall, and 96.59% F1-score, outperforming the DenseNet201 + Bayesian-optimized SVM baseline (94.60%, 94.40%, 94.40%, and 94.40%, respectively). This improvement corresponds to a 2.98% accuracy gain, a 55% reduction in error rate, an 86% reduction in model parameters (20.31M to 2.75M), and an 85% reduction in model size (81 MB to 12 MB). These results indicate that the proposed framework provides a compact and efficient solution with strong potential for deployment in resource-constrained agricultural environments.

An Attention-Enhanced CNN–RBF Framework for Network Intrusion Detection in Imbalanced Traffic

Fabrice Kabura, Thierry Nsabimana — Mon, 23 Feb 2026 00:00:00 +0000

The increasing complexity and scale of modern network traffic driven by IoT and cloud-based infrastructures have made accurate intrusion detection a critical challenge. Conventional network intrusion detection systems (NIDS) and many deep learning–based approaches struggle to reliably detect minority and stealthy attacks due to severe class imbalance and limited discrimination of subtle traffic patterns. To address these limitations, this study proposes a hybrid CNN–RBF–Attention framework for network intrusion detection. The proposed model integrates three complementary components: (i) a convolutional neural network for hierarchical feature extraction from network flow data, (ii) a radial basis function (RBF) network for localized nonlinear classification using prototype-based decision regions, and (iii) an attention mechanism that adaptively weights RBF activations to emphasize discriminative traffic patterns. SMOTE is applied exclusively to the training data to mitigate class imbalance. The framework is evaluated on the widely used CICIDS2017 and CICIDS2018 benchmark datasets in both binary and multiclass settings, using recall, precision, F1-score, confusion matrices, and ROC analysis. Experimental results demonstrate that the proposed hybrid model consistently outperforms standalone CNN and RBF baselines, particularly in terms of recall and F1-score. On the CICIDS2018 dataset, the model achieves 99.81% accuracy and 99.81% F1-score in binary classification, and 99.54% accuracy and 99.54% F1-score in multiclass classification. On CICIDS2017, it achieves 98.12% accuracy and 98.12% F1-score in binary classification, and 98.92% accuracy and 98.92% F1-score in multiclass classification. Confusion matrix and ROC analyses further show strong class separability and reliable performance in low–false-positive-rate regions, which is critical for real-world IDS deployment. These results confirm that combining deep hierarchical feature learning, localized prototype-based classification, and attention-guided refinement yields a robust, operationally reliable intrusion detection framework for highly imbalanced network environments.

A Graph-Augmented Isolation Forest Using Node2Vec and GraphSAGE for Mobile User Behavior Anomaly Detection

Mon, 23 Feb 2026 00:00:00 +0000

This study presents a Graph-Augmented Isolation Forest (GAIF), an unsupervised anomaly-detection framework for analyzing mobile user behavior. The proposed framework represents users and behavioral attributes as a user–feature bipartite graph, enabling the capture of relational dependencies that are not explicitly modeled in conventional vector-based approaches. Low-dimensional user representations are learned through Node2Vec and Graph Sample and Aggregate (GraphSAGE), and the resulting embeddings are subsequently processed by an Isolation Forest to produce anomaly scores. Experiments are conducted on a Mobile Device Usage and User Behavior dataset comprising 700 user profiles derived from application-level behavioral indicators. The dataset is treated as a behavioral abstraction rather than as a malware classification benchmark. A consistent 80:20 stratified train–test split is employed, with all learning-capable operations restricted to the training data to mitigate information leakage. Detection performance is evaluated post hoc using precision, recall, F1-score, and area under the curve (AUC) metrics. Under the evaluated setting, GAIF achieves an F1-score of 0.94 and an AUC of 0.97, demonstrating improved anomaly detection effectiveness relative to representative unsupervised baseline methods. These results are obtained on a static, proxy dataset and should not be interpreted as evidence of real-time deployment capability. Model interpretability is supported through post-hoc Uniform Manifold Approximation and Projection (UMAP) visualizations of the learned embeddings, providing structural insights into anomalous user behavior. Overall, the findings indicate that integrating graph-based representation learning with isolation-based anomaly scoring constitutes a computationally efficient approach for unsupervised mobile user behavior anomaly detection within the scope of this study.

Harnessing Artificial Intelligence for Early Disease Detection: Opportunities and Challenges in Modern Healthcare

Achile Solomon Egbunu, Akindele Michael Okedoye — Mon, 23 Feb 2026 00:00:00 +0000

Artificial Intelligence (AI) is increasingly recognized as a transformative enabler of early disease detection, with the potential to improve diagnostic accuracy, support predictive risk stratification, and advance preventive healthcare. Despite rapid methodological progress, many existing reviews remain performance-centric, offering limited insight into generalizability, ethical governance, and real-world implementation constraints. This paper presents a narrative and integrative review with an adoption-focused, translational perspective, synthesizing recent developments in AI-driven early disease detection across oncology, cardiology, neurology, and infectious disease surveillance. Drawing on peer-reviewed literature published primarily between 2016 and 2025, the review examines reported performance gains alongside persistent limitations related to data heterogeneity, population bias, explainability, and regulatory fragmentation. Through cross-sectional synthesis, we identify three recurring gaps in prior reviews: (i) overgeneralization of AI’s diagnostic superiority, (ii) insufficient consideration of ethical and legal accountability, and (iii) a lack of actionable guidance for scalable clinical implementation. Integrating technical, ethical, and policy dimensions into a unified conceptual framework, this review demonstrates that while AI systems can consistently enhance diagnostic accuracy and early risk stratification in well-defined tasks, sustained clinical adoption depends on aligning technical performance with governance readiness, interpretability, and workflow integration. The analysis further highlights how implementation mechanisms—such as explainable AI, continuous post-deployment monitoring, and clinician-centered deployment strategies—mediate the translation of algorithmic innovation into real-world healthcare impact. Overall, this review provides a critical reference for researchers, clinicians, and policymakers seeking to translate AI innovation into safe, equitable, and trustworthy clinical practice.

Immersive Interventions for Dementia: A Narrative Review of Virtual Reality's Role in Therapy, Well-Being, and Future Care Models

Prathibha Samarasekara, Kasun Karunanayaka, Sanjani Gunathilaka — Mon, 23 Feb 2026 00:00:00 +0000

Dementia is a progressive neurocognitive disorder often accompanied by behavioral and psychological symptoms such as agitation, anxiety, and depression. Pharmacological treatments provide only modest benefits while introducing significant risks, which highlights the need for safer, non-pharmacological alternatives. This literature review examines the role of virtual reality in dementia care, with a focus on its integration with therapies such as music, reminiscence, sensory stimulation, and cognitive training. Evidence from prior research suggests that virtual reality can enhance cognitive functions, reduce symptoms, and improve emotional well-being while also strengthening patient–caregiver interaction. However, challenges related to usability, accessibility, cost, and long-term effectiveness continue to limit adoption. Gaps in research, including limited cultural diversity, inconsistent reporting of intervention design, and a lack of large-scale longitudinal trials, emphasize the need for future work exploring cross-cultural feasibility and AI-driven personalization. Overall, virtual reality represents a promising and evolving non-pharmacological intervention that has the potential to transform dementia care by improving quality of life and reducing reliance on medication.