Journal of Computing Theories and Applications

YOLOv9s with Region-Dispersion Channel Spatial Attention for Robust Chili Leaf Disease Detection

2026-05-15T14:54:22+00:00

Abstract: Detecting chili leaf diseases remains challenging due to the non-uniform manifestation of symptoms, local discoloration, small lesion regions, and visual similarity between disease patterns and natural leaf background variations. Although YOLO-based detectors provide favorable computational efficiency, lightweight variants often struggle to distinguish subtle lesion characteristics, while conventional attention mechanisms such as CBAM primarily rely on global feature aggregation and may overlook regional activation variability. To address these limitations, this study proposes a YOLOv9s-based detection framework integrated with a Region-Dispersion Channel Spatial Attention (RDCSA) module. The proposed module incorporates regional dispersion statistics, namely mean, standard deviation, and range, as channel descriptors to capture inter-region feature variability before applying spatial attention refinement. Experiments were conducted on the COLD dataset containing 532 original images from five chili leaf condition categories using a split-before-augmentation protocol to ensure objective evaluation. RDCSA was integrated at the P5 feature level and evaluated through attention placement analysis, component-wise ablation, sensitivity analysis, stability assessment, and comparison with modern attention mechanisms. The proposed YOLOv9s + RDCSA model achieved an mAP@50 of 0.894, mAP@50–95 of 0.773, precision of 0.858, recall of 0.861, and an F1-score of 0.859 with only a marginal increase in model parameters. The results suggest that regional dispersion-based attention improves feature discrimination while preserving computational efficiency, particularly for disease symptoms characterized by heterogeneous spatial patterns. Nevertheless, performance remains influenced by visually ambiguous symptom categories, indicating that further validation across multiple datasets and field conditions is required. Overall, the proposed RDCSA module enhances detection capability without substantially increasing computational overhead, making it a promising attention mechanism for lightweight plant disease detection systems.

A Systematic Literature Review of Robustness-Aware Batik Motif Classification: Acquisition Variability, Feature Representation, and Learning Models

2026-04-30T05:38:22+00:00

Batik motif classification has attracted growing attention in visual computing due to its role in cultural heritage preservation, textile informatics, museum documentation, and automated cataloging. Although many studies report high classification accuracy, robustness under real-world acquisition conditions remains insufficiently understood. Batik images are frequently affected by illumination variation, blur, folds, watermark overlays, wearable deformation, scale inconsistency, and background clutter, creating challenges that extend beyond conventional image-noise assumptions. Existing studies largely focus on improving classification performance, while the interactions among acquisition variability, feature representation, evaluation practice, and deployment constraints remain fragmented. This systematic literature review addresses this gap by synthesizing batik classification research through a robustness-aware perspective. Using query expansion, backward and forward citation chaining, relevance screening, and thematic coding, 116 candidate records were identified, resulting in 50 highly relevant studies for detailed analysis. The review reveals that robustness is shaped less by denoising alone than by the combined effects of acquisition conditions, representation design, evaluation realism, and deployment context. Handcrafted descriptors remain competitive for small datasets and structured motifs due to their data efficiency and interpretability, whereas deep learning models achieve the highest reported accuracy when supported by sufficient data diversity and realistic augmentation. Hybrid representations emerge as the most consistently balanced approach, combining local texture stability with higher-level abstraction across heterogeneous acquisition settings. The review further identifies recurring robustness failure patterns, including background dependency, illumination instability, motif-scale inconsistency, wearable deformation, and source-shift vulnerability. Based on these findings, a robustness-oriented research agenda is proposed, emphasizing cross-acquisition evaluation, representation-stability analysis, batik-specific robustness benchmarks, acquisition-aware augmentation, and deployable lightweight or hybrid architectures. The study contributes a domain-specific synthesis that reframes batik motif classification from an accuracy-centric task toward a robustness-aware visual recognition problem.

A Systematic Review of Agentic AI in Healthcare: An Evidence-Informed Seven-Principle Framework

2026-05-30T15:12:10+00:00

Agentic artificial intelligence (AI) systems capable of autonomous goal-directed behavior, multi-step planning, tool use, multi-agent coordination, and iterative self-correction represent a transition from passive clinical AI tools toward systems that can participate in complex healthcare workflows. However, empirical evidence remains fragmented across clinical decision support, patient monitoring, and administrative applications, and no systematic synthesis has evaluated which agentic principles have been technically demonstrated and which have accumulated sufficient evidence to support responsible clinical deployment. We conducted a PRISMA-informed systematic review of peer-reviewed empirical studies published between January 2025 and April 2026. Searches across five bibliographic databases and Google Scholar, supplemented by citation tracking, identified 443 unique records for screening, of which 25 met the predefined PICOS and quality appraisal criteria. Evidence was synthesized using an evidence-informed seven-principle framework derived from the integration of agentic AI, clinical AI, and healthcare governance literature. This framework provides a structured lens for examining how agentic principles are evaluated individually and in combination, enabling a deployment-readiness perspective that extends beyond capability-focused assessments alone. The evidence base was concentrated on technical capability principles, whereas human oversight, safety, compliance, and equity-related evaluation received comparatively limited attention. Most studies remained at the laboratory, benchmark, or proof-of-concept stage, and none reported demographic-stratified performance outcomes. Overall, the findings suggest a structural asymmetry in agentic healthcare AI: empirical research is advancing agentic capabilities more rapidly than it is generating evidence for the oversight, safety, equity, and governance mechanisms required for responsible clinical translation.

Sentence-Level Sentiment Analysis of Indonesian App Reviews Using IndoBERTweet

2026-05-21T01:17:50+00:00

Document-level sentiment analysis assigns a single polarity label to an entire review, often obscuring opinion diversity within multi-sentence submissions. This limitation is particularly evident in reviews of multi-service platforms, where users frequently express heterogeneous opinions toward different aspects of the platform in the same review. To address this challenge, this study proposes a sentence-level sentiment analysis framework for Indonesian Gojek app reviews collected from the Google Play Store. The proposed framework introduces a two-stage segmentation strategy that combines punctuation-aware rules with conjunction-aware splitting based on coordinating and adversative conjunctions (e.g., tapi [but], padahal [even though]) to identify opinion boundaries and decompose mixed-sentiment reviews into independently classifiable sentence units. A total of 14,730 raw reviews collected between May and July 2025 were subjected to data cleaning and quality filtering, resulting in 7,187 valid reviews that were further segmented into 14,187 sentence-level instances. Each instance was manually annotated by three annotators using a four-class labeling scheme consisting of app-positive, app-negative, app-neutral, and service categories. Sentiment-level inter-annotator agreement, computed on the subset of instances unanimously categorized as app-related by all three annotators (n = 4,384), achieved substantial agreement (Fleiss' = 0.636). Hyperparameter optimization was conducted using Optuna with the Tree-structured Parzen Estimator (TPE) sampler across four experimental scenarios. The best performance was achieved by IndoBERTweet under Stratified K-Fold evaluation, attaining an accuracy of 0.751 and a macro F1-score of 0.729, outperforming all IndoBERT configurations. The results demonstrate the effectiveness of domain-adaptive pre-training on informal Indonesian text and highlight the value of conjunction-aware segmentation for preserving fine-grained opinion structures in mixed-sentiment reviews. These findings suggest that domain-aligned language representations provide a practical and effective solution for sentence-level sentiment analysis of Indonesian app reviews.

A Composite Centrality Framework for Evacuation Planning in Meso-Scale Spatial Networks with Semi-Structured Connectivity

2026-04-15T18:18:46+00:00

Evacuation planning in spatial networks requires the identification of critical nodes that maintain connectivity, accessibility, and flow distribution during emergency situations. Existing approaches often rely on individual centrality measures, which capture only a single structural dimension of node importance and may therefore produce incomplete or biased prioritization. To address this limitation, this study proposes a Composite Centrality Framework for identifying critical nodes in meso-scale spatial networks with semi-structured connectivity. The network is modeled as a weighted undirected graph, and Degree, Betweenness, and Closeness Centrality are integrated into a unified composite index to capture complementary structural roles. The framework is implemented in MATLAB and evaluated using a real-world campus spatial network consisting of 30 nodes and a synthetic network comprising 16 nodes with comparable structural characteristics. The results reveal a highly uneven distribution of node importance, with a small set of structurally dominant nodes consistently identified across both networks. In the campus network, node P1 achieves the highest composite centrality score (0.2195) and ranks first across the individual centrality measures, indicating its dominant role in maintaining network connectivity, accessibility, and flow distribution. Quantitative evaluation demonstrates strong agreement between the composite ranking and the individual measures, with Spearman rank correlation coefficients of 0.94, 0.89, and 0.91 for Degree, Betweenness, and Closeness Centrality, respectively. However, only one node (P1) appears simultaneously in the top five of all rankings, highlighting the complementary nature of the individual centrality measures and supporting the need for multi-criteria integration. Sensitivity analysis across three weighting scenarios yields rank correlations exceeding 0.97, confirming ranking stability and methodological robustness. Overall, the proposed framework provides a balanced and reliable approach for identifying critical nodes and demonstrates potential applicability to evacuation planning and spatial network analysis in semi-structured environments.

AN-RPL: Infrastructure-Assisted RPL Enhancement via Distributed Anchor Nodes for Mobile IoT Networks

2026-05-21T02:02:57+00:00

The Internet of Things (IoT) has attracted significant attention from the research community due to its wide range of applications. However, the limited energy, processing capability, storage, and communication capacity of IoT devices require routing solutions that are both lightweight and efficient. To address these constraints, the IPv6 Routing Protocol for Low-Power and Lossy Networks (RPL) was introduced in 2012 as a routing protocol specifically designed for resource-constrained IoT environments. Although RPL performs reliably in static deployments, its performance degrades considerably in mobile environments because of frequent topology changes, slow Trickle timer convergence, and excessive parent churn. This paper proposes Anchor-Node RPL (AN-RPL), an infrastructure-assisted enhancement of RPL that strategically deploys distributed fixed anchor nodes as stable DODAG roots while requiring only minimal firmware modification on mobile sensor nodes, namely a single anchor-flag check during parent selection. Simulation experiments conducted in Cooja using both OF0 and MRHOF objective functions across four scenarios (static, mobile with one, two, and four anchor nodes) demonstrate that AN-RPL with four anchor nodes improves the Data Delivery Ratio (DDR) by up to 30.6 percentage points, reduces the average hop count by up to 51.2%, lowers parent churn by up to 89.5%, and decreases average energy consumption by up to 14.8% compared with conventional single-root mobile RPL. These results demonstrate that infrastructure-assisted anchor deployment provides an effective and practical approach for improving routing reliability and efficiency in mobile RPL-based IoT networks.

Secure Lightweight Face Authentication with MTCNN and LightCNN for Digital Financial Services

2026-06-05T15:59:55+00:00

Mobile face authentication for digital financial services must simultaneously satisfy recognition accuracy, computational efficiency, and biometric security under resource-constrained deployment conditions. This study proposes and evaluates a lightweight face-authentication framework that integrates detector–recognition pipeline optimization, protected biometric-template transformation, and blockchain-backed integrity support. Five face detection–recognition pipelines were systematically evaluated using a shared LightCNN-29v2 backbone fine-tuned on the Indonesian Muslim Student Face Dataset (IMSFD), with Mahalanobis-based Distance-Based Encryption (DBE) providing protected template matching and blockchain hash anchoring serving as an architectural integrity layer. Experiments on 3,660 images from 68 identities demonstrate that the MTCNN + LightCNN pipeline achieves the most favorable in-domain performance, reaching 94.95% accuracy, a ROC-AUC of 0.9970, an F1-score of 0.95, and successful processing of 3,546 out of 3,660 test images with an overall model size of approximately 5 MB. Applying Mahalanobis-based DBE further improves verification performance on IMSFD, increasing accuracy to 96.88% while reducing the False Acceptance Rate (FAR) from 0.83% to 0.18% and the False Rejection Rate (FRR) from 16.44% to 10.12%. Cross-dataset evaluation on LFW, CFP-FF, CFP-FP, AgeDB-30, CALFW, and CPLFW indicates that the proposed framework generalizes well to frontal-domain benchmarks but exhibits expected performance degradation under cross-pose and cross-age conditions due to distribution shift. Overall, the results demonstrate that detector selection is the dominant factor influencing end-to-end verification performance, while domain-specific fine-tuning and protected template matching are essential for secure and practical deployment in mobile financial authentication systems.

Cross-Domain Faithfulness Evaluation of SHAP and Attention-Based Explanations in Transformer NLP Models

2026-06-05T15:22:41+00:00

Transformer-based models such as BERT, RoBERTa, DistilBERT, and DeBERTa have achieved remarkable performance across a wide range of natural language processing (NLP) tasks. However, their decision-making processes remain difficult to interpret, particularly in high-risk applications such as hate speech detection, where unreliable explanations may undermine model transparency, trust, and accountability. This study investigates whether explainability methods remain faithful and stable under domain shift in transformer-based text classification. Four transformer architectures were fine-tuned and evaluated on two linguistically distinct datasets: IMDb Movie Reviews and Hate Speech Offensive. Model performance and explanation quality were assessed using classification accuracy, macro F1-score, top-k token-removal faithfulness analysis, and cross-domain Spearman rank correlation. Experimental results show that DeBERTa achieved the highest classification performance, reaching accuracies of 95.6% on IMDb and 91.3% on Hate Speech. Across all evaluated models and datasets, SHAP consistently produced higher faithfulness scores than attention-based explanations. Cross-domain analysis further revealed reduced agreement between SHAP and attention-based explanations under domain shift, indicating lower explanation consistency across linguistically distinct domains. Qualitative error analysis further showed that implicit sentiment, sarcasm, and domain-specific slang remain major sources of prediction errors. Overall, the results demonstrate that superior predictive performance does not necessarily correspond to higher explanation faithfulness or stronger cross-domain stability. These findings highlight the importance of jointly evaluating predictive performance, explanation faithfulness, and explanation robustness when developing trustworthy transformer-based NLP systems.

A Side-Channel-Aware Cryptographic Framework for Secure Interactive, Embedded, IoT, and Edge Communication Systems

2026-06-29T08:22:07+00:00

Secure interactive, embedded, and edge communication systems are often deployed in physically exposed environments where algorithmically secure ciphers may exhibit implementation-specific leakage. This paper presents a simulation-based evaluation of a configurable hybrid AES-Blowfish framework comprising AES-Hybridization, Combined Blowfish Trilogy, and cascaded Blowfish-to-AES modes. The AES-Hybridization path combines AES-256-CBC, Argon2id-derived whitening material, HMAC-based integrity binding, plaintext masking, and modeled randomized hiding activity. The hiding activity is represented solely within the leakage simulation and does not modify the plaintext or ciphertext. The Blowfish path employs session-dependent P-array randomization, dynamic S-box initialization, and a three-stage Feistel-like structure. The framework is evaluated using representative IoT/edge workload proxies, component-level ablation studies, modeled leakage assessment, non-ideal leakage scenarios, parameter-sensitivity analysis, and a software-level performance model. Compared with the simulated baseline configurations, the proposed modes reduce modeled TVLA, CPA, and DPA distinguishability, with AES-Hybridization providing the most balanced security-performance trade-off and the cascaded mode achieving the lowest modeled distinguishability at the highest computational cost. All findings are derived exclusively from simulation; no validation using physical power traces, electromagnetic traces, FPGA implementations, or microcontroller platforms is claimed.

Retinal Vessel Segmentation via Morphological Refinement and Adaptive Late Fusion

2026-05-01T02:03:51+00:00

Retinal vessel segmentation is fundamental for quantitative retinal vascular analysis; however, accurate delineation remains challenging because of nonuniform illumination, low-contrast capillaries, pathological lesions, and variations in image acquisition. This study presents a rule-based unsupervised retinal vessel segmentation framework based on decision-level adaptive late fusion, in which six complementary local adaptive thresholding methods are integrated and subsequently refined through luminance validation, hysteresis reconstruction, component cleanup, elongation filtering, and boundary smoothing. In this context, unsupervised denotes that no statistical segmentation model is trained using manual vessel annotations; instead, fixed parameters and fusion weights are determined from a small development subset, while all evaluation images remain unseen during method development. The proposed framework was evaluated on 135 independent retinal fundus images from the DRIVE, STARE, CHASE_DB1, HRF, and LES-AV datasets using an eroded field-of-view protocol. It achieved an image-weighted Dice coefficient of 0.7104 (95% bootstrap confidence interval: 0.7005–0.7205), an IoU of 0.5539, a sensitivity of 0.7544, a specificity of 0.9597, a balanced accuracy of 0.8570, and a clDice score of 0.7395. Compared with fixed majority voting, the proposed adaptive late fusion strategy significantly improved segmentation performance in 128 of 135 test images, yielding a mean Dice improvement of 0.0270 (paired Wilcoxon, Holm-adjusted p = 7.67 × 10⁻²¹). Although segmentation performance remains limited for complex pathological images, particularly in the HRF dataset, the proposed framework demonstrates consistent cross-dataset performance, transparent decision-making, and strong reproducibility, providing an effective alternative when interpretable, training-free retinal vessel segmentation is required.

Beyond Binary Fraud Detection: Amount-Aware Operational Ranking for Transaction Risk Prioritization

2026-06-09T09:07:05+00:00

Fraud detection in digital financial transactions is traditionally formulated as a binary classification problem, although real-world fraud investigation requires analysts to prioritize a limited number of suspicious transactions according to operational risk and potential financial impact. This study reformulates fraud detection as an amount-aware operational ranking problem for fraud-risk prioritization. Transactions are organized into time-window query groups, and fraudulent transactions are assigned graded relevance based on training-only transaction-amount quartiles, enabling the ranking objective to distinguish low- and high-severity fraud without relying on proprietary cost matrices. The proposed formulation is implemented using a representative Learning-to-Rank framework based on LambdaMART, while an out-of-fold XGBoost risk score is incorporated as an auxiliary feature to refine the ranking representation rather than serve as the primary contribution. Experiments conducted on a public credit-card fraud dataset using chronological validation and future-holdout testing demonstrate that amount-aware relevance consistently improves severity-aware top-rank ordering compared with conventional binary relevance. The proposed HybridLTR_amount model significantly outperforms XGBClassifier and PureLTR_binary in terms of all-query NDCG@10, whereas its performance is not statistically different from PureLTR_amount, indicating that the primary empirical improvement is attributable to the amount-aware ranking formulation rather than the auxiliary hybrid component. Additional operational analyses show that high-risk transactions and fraudulent financial losses are concentrated within a compact top-ranked segment, while budget-oriented evaluation demonstrates the practical value of the proposed formulation under limited analyst review capacity. These findings establish amount-aware operational ranking as an effective formulation-centric framework for operational fraud-risk prioritization rather than as a new classification algorithm.

Block-wise Authenticated DNA-based Image Encryption with Tamper Localization using HKDF-Derived Keys

2026-06-16T00:02:27+00:00

Confidentiality alone cannot establish whether a medical, forensic, cloud-stored, or remotely sensed image has been modified during transmission, while a single global authentication verdict cannot identify the affected region. This paper presents a block-wise authenticated DNA-based image encryption framework that integrates a DNA-chaos confidentiality core with pre-decryption integrity verification and spatial tamper localization. An authenticated ephemeral X25519 key exchange establishes a session secret, which is expanded by HKDF-SHA256 into four transcript-bound, domain-separated subkeys for permutation, DNA operations, diffusion, and authentication. Chaotic seeds are derived from a canonical plaintext hash and block coordinates, preserving strong plaintext differential sensitivity while confining post-encryption modifications to the affected blocks. Ciphertext integrity is enforced using a block-wise encrypt-then-MAC construction with 128-bit truncated HMAC-SHA256 tags that authenticate both the canonical header and each ciphertext block. A reduction-based security argument shows that the authentication layer provides ciphertext integrity and conditionally upgrades an IND-CPA encryption core to IND-CCA security under standard HKDF and HMAC assumptions, while the confidentiality claim remains explicitly conditional on the encryption core. Experiments on 30 tuberculosis chest radiographs across five independent sessions achieved a ciphertext entropy of 7.9972, near-zero adjacent-pixel correlation, 99.61% NPCR, and 33.47% UACI. Across five representative tampering attacks, the proposed framework achieved an observed 100% block-level true-positive rate, 0% false-positive rate, and required only 2.847 ± 0.258 ms for block-wise authentication with 6.25% tag overhead using the default 16×16 block configuration. These results demonstrate that the proposed framework effectively combines statistical confidentiality, modern cryptographic key management, and reliable block-level tamper localization within a unified authenticated image encryption architecture.