Computer and Information Science, Issue: Vol.19, No.2

Toward Epistemic Stability: Engineering Consistent Procedures for Industrial LLM Hallucination Reduction

Tue, 26 May 2026 07:53:28 +0000

Hallucinations in large language models (LLMs) are outputs that are syntactically coherent but factually incorrect or contextually inconsistent. They are persistent obstacles in high-stakes industrial settings such as engineering design, enterprise resource planning, and IoT telemetry platforms. We present and compare five prompt engineering strategies intended to reduce the variance of model outputs and move toward repeatable, grounded results without modifying model weights or creating complex validation models. These methods include: (M1) Iterative Similarity Convergence, (M2) Decomposed Model-Agnostic Prompting, (M3) Single-Task Agent Specialization, (M4) Enhanced Data Registry, and (M5) Domain Glossary Injection. Each method is evaluated against an internal baseline using an LLM-as-Judge framework over 100 repeated runs per method (same fixed task prompt, stochastic decoding at τ = 0.7. Under this evaluation setup, M4 (Enhanced Data Registry) received “Better” verdicts in all 100 trials; M3 and M5 reached 80% and 77% respectively; M1 reached 75%; and M2 was net negative at 34% when compared to single shot prompting with a modern foundation model. We then developed enhanced version 2 (v2) implementations and assessed them on a 10-trial verification batch; M2 recovered from 34% to 80%, the largest gain among the four revised methods. We discuss how these strategies help overcome the non-deterministic nature of LLM results for industrial procedures, even when absolute correctness cannot be guaranteed. We provide pseudocode, verbatim prompts, and batch logs to support independent assessment.

AI-Based Data Quality Management and Cleaning in Intelligent Databases

Mon, 15 Jun 2026 10:38:35 +0000

The use of large datasets for predictive modeling in intelligent databases is frequently hindered by structural deficiencies, most notably missing records. Conventional imputation methods, such as mean or median substitution, distort statistical variance and fail to capture multivariate relationships among features. This paper presents a dynamic, AI-based data quality management framework that autonomously identifies and corrects structural anomalies prior to predictive modeling. Utilizing Classification and Regression Trees (CART), the system intelligently imputes missing categorical and numerical values by learning localized patterns from the observed portion of the dataset. To validate the integrity of the repaired database, the framework was evaluated on a credit risk dataset using a Tri-Ensemble of gradient boosting algorithms (XGBoost, LightGBM, and CatBoost), achieving an Area Under the ROC Curve (AUC) score of 0.9393. These results demonstrate that predictive, tree-based data imputation preserves the statistical distribution of the original data and substantially enhances the accuracy of downstream machine learning tasks. The proposed framework offers a scalable, automated solution for data preprocessing in distributed and intelligent database environments.

A Comprehensive Taxonomy of Cryptographic Attacks on the Advanced Encryption Standard (AES-128)

Mon, 15 Jun 2026 10:45:02 +0000

The 128-bit Advanced Encryption Standard is the only block cipher that is known to be widely deployed among the other block ciphers, in relation to disk encryption, transport layer security, payment gateways and secure communications. Despite its ubiquitous nature, there is fragmented literature covering the broad spectrum of attacks on this scheme. Algebraic and differential attacks are covered in a theoretical nature, side channel attacks in either embedded systems or journals of hardware security and statistical tests in implementation evaluations standards, there are fragmented and rarely found in one literature review document. This study consolidates a unified document spanning sixteen principal attacks across twelve categories, such as differential cryptanalysis, brute force attacks, linear cryptanalysis, algebraic attacks, biclique attacks, slide attacks, cube attacks, boomerang attacks, statistical attacks, cache timing side channel attacks, impossible differential attacks, integral (Square) attacks, meet-in-the-middle attacks, related-key attacks and power analysis side-channel attacks. Each attack model has a formal mathematical expression, the precise data and time complexity as applied to AES, the nature of AES resistance mechanism and the residual practical risks where possible. A comparative evaluation against nine previous works shows that there exists no detailed and in-depth work as demonstrated in this research. One of the main objectives of this study is to produce a unified document that serves as a cryptanalysis guide and a resource for researchers working on developing new variants of AES to counter existing attacks. The guide further contains details on the structural limitations of AES and its modified variants.

Does Explainability Translate to Better Decisions? A Controlled Behavioural Experiment on Program Manager Decision Quality with and without Explainable Artificial Intelligence

Mon, 13 Jul 2026 07:31:16 +0000

Explainable Artificial Intelligence (XAI) is rapidly being integrated into project and program management decision-support tools, driven by the widely held assumption that model transparency improves human decision quality. This assumption has not been subjected to rigorous empirical testing in program management contexts. This paper presents the design, theoretical framework, experimental methodology, and pilot study findings of a controlled behavioural experiment evaluating whether program managers who receive XAI-augmented risk predictions make measurably better decisions than those receiving prediction-only outputs or no AI assistance. Three experimental conditions are defined: unaided judgment, prediction-only AI output, and XAI-augmented AI output with SHAP-derived feature explanations. Four dependent variables are measured: decision accuracy, intervention targeting precision, decision latency, and stakeholder communication quality. A pilot study conducted with 18 practitioners provides preliminary support for the primary hypotheses, with moderate-to-large effect size estimates confirming adequate statistical power for the full study. The framework is grounded in Dual Process Theory, Bounded Rationality, and the Technology Acceptance Model. The study addresses a critical empirical void at the intersection of explainable AI and program governance and is positioned to produce the first empirically validated evidence base on the decision-level value of XAI in IT program risk management.

Effectiveness of National ID as a Unique Identifier for Customer Onboarding in Financial Services in Bangladesh: A Survey

Mon, 13 Jul 2026 23:37:44 +0000

National identity systems lie at the heart of every government’s identity management infrastructure. In Bangladesh, the National Identification Number (NID) has emerged as the primary credential for citizen identification, yet its integration with the financial sector—particularly for electronic Know Your Customer (e-KYC) compliance—remains fragmented, inconsistent, and vulnerable. This study investigates the effectiveness of the NID as a single authentication identifier for customer onboarding in Bangladesh’s banking and non-banking financial institution (NBFI) sector. We conducted semi-structured interviews with ICT heads and technically responsible personnel across 61 banks and 35 NBFIs, supplemented by a structured questionnaire survey covering six key dimensions: single authentication ID usage, e-KYC service operations, onboarding challenges, personally identifiable information (PII) handling, identity theft exposure, and agent banking practices. A mixed-methods analytical approach was applied, combining quantitative survey responses with qualitative thematic coding. Our findings reveal sixteen distinct vulnerabilities and operational gaps grouped into three categories: identity theft, privacy vulnerabilities, and technical vulnerabilities. Notably, the majority of institutions rely on manual, paper-dependent verification processes; the Bangladesh Election Commission (BEC) API constitutes a single point of failure for onboarding verification across hundreds of organizations; and a significant proportion of banks retain NID data beyond its intended purpose. Despite these challenges, we demonstrate that a Federated Identity Management (FIM) system built on the existing NID infrastructure is technically feasible and would substantially reduce administrative burden, improve security, and enable seamless cross-sector identity federation.

A Chaos Theory Perspective on Expert Decision-Making in Cyber Incident Response

Mon, 13 Jul 2026 07:44:38 +0000

This manuscript is an abridged version of a doctoral dissertation exploring the critical human dimensions of cyber incident response. Modern technology environments and volatile threat landscapes generate unpredictable cyber crises that strain traditional, rigid defense playbooks and demand unprecedented responder adaptability. To understand the critical human elements required to navigate these crises, this research applies chaos theory to analyze how expert cyber incident responders (CIRs) manage nonlinear, high-pressure events. Employing a qualitative phenomenological design, the study investigates the lived experience of expert CIRs, focusing on the cognitive and affective processes that drive decision-making during systemic failures. The findings reveal that successful incident resolution relies not on strict adherence to linear protocols, but on the rapid, iterative generation of novel solutions. Responders intentionally leverage course-correction as a core functional mechanism, transforming unexpected adversarial inputs into actionable intelligence via high-velocity feedback loops. By adapting chaos theory to cybersecurity operations, this study provides a theoretical model explaining how adaptive behaviors emerge under pressure. Ultimately, these insights offer a foundation for developing dynamic training and policy frameworks that cultivate cognitive flexibility and bolster organizational and national cyber resilience.

The Digital Thread: Engineering Purpose, Limitation, and Consent in Disparate Cloud Ecosystems

Tue, 04 Aug 2026 05:20:28 +0000

As organizations transition to hyper-connected, multi-cloud architectures, the mobility of data has far outpaced the technical frameworks meant to govern it. Regulatory mandates GDPR, CCPA, and CPRA articulate purpose limitation and informed consent as foundational principles, yet a persistent gap exists between what these laws require and what distributed systems actually enforce. In cloud-native environments, the metadata that records user consent and allowable use is routinely lost or overwritten as data crosses microservice boundaries, IoT pipelines, and third-party APIs. This article examines the architectural roots of that failure through three documented enforcement actions: the 2026 FTC settlement with General Motors over its Smart Driver telematics program, Tesla's internal camera-data misuse, and the Toyota-Progressive Insurance opt-out breakdown. Each case traces the same systemic flaw, consent does not travel with the data it authorizes. In response, this article proposes a Sticky Governance framework built on cryptographically signed consent tokens, service mesh enforcement, and dynamic resource provisioning, together with a revised audit methodology appropriate for information systems professionals operating in automated, high-velocity data environments.

Analysis of three documented enforcement actions reveals a consistent pattern: purpose-scope metadata was absent from inter-organizational API payloads in all cases examined, enabling downstream data misuse that produced a 2026 FTC consent order, internal privacy violations affecting video footage from tens of thousands of enrolled vehicles, and insurance opt-out failures persisting across at least one underwriting cycle. The Sticky Governance Framework operationalizes cryptographic consent propagation and automated purpose-matching at the service-mesh enforcement layer, directly addressing the architectural gap through which governance context is stripped in transit. Policy recommendations include mandatory consent-token schema standards for cross-organizational API ecosystems, governance-metadata preservation clauses in cloud service contracts, and audit frameworks grounded in cryptographic evidence rather than document-review sampling.

HTTP Anomaly Detection Using Character-Level 3-Gram TF-IDF and Hybrid TF-IDF–Hashing Representations: A Comparative Study of Machine Learning Models

Thu, 03 Sep 2026 09:41:27 +0000

Web applications are frequent targets of cyberattacks such as SQL injection, cross-site scripting, and command injection, creating a need for accurate and computationally efficient anomaly detection systems. While recent approaches increasingly rely on deep learning and transformer-based architectures, these methods often require substantial computational resources and deployment complexity. This study presents a comparative evaluation of character-level feature representations and machine learning models for HTTP anomaly detection. Two feature encoding strategies, namely character-level 3-gram TF-IDF and hybrid TF-IDF–hashing representations, are investigated using Random Forest, LightGBM, multilayer perceptron (MLP), and fusion-based classifiers on the CSIC 2010 benchmark dataset. All models are trained and evaluated under consistent experimental settings to enable a fair comparison of feature representations and classification approaches. Experimental results show that the hybrid TF-IDF–hashing representation combined with LightGBM achieves the best performance, attaining 99.60% accuracy, precision, recall, and F1-score, outperforming all other evaluated models. Furthermore, comparisons with recently reported deep learning and transformer-based approaches indicate that lightweight machine learning methods can achieve competitive or superior performance while maintaining lower computational complexity and deployment overhead. The findings provide practical insights into the trade-offs between feature representations and classification models for HTTP anomaly detection.

Evidence-Graph Dual-Store Multimodal RAG for Brain Tumor Evaluation from MRI

Mon, 21 Sep 2026 11:01:15 +0000

This study proposes an evidence-graph dual-store multimodal retrieval-augmented generation (MM-RAG) framework that combines MRI-derived visual evidence with textual evidence retrieved from similar historical cases and from a curated clinical knowledge repository. The framework couples a case-memory index with a knowledge index, and an agreement-weighted retrieval score selects evidence that is consistent across the two stores. Retrieved items are organised as an evidence graph rather than concatenated, so that supportive signals propagate while unrelated context is suppressed. Experiments use publicly available multiparametric MRI collections and are restricted to three classification tasks: tumor detection, tumor subtype classification, and tumor grade prediction. The framework is compared against three baselines: a vision-only 3D ViT/CNN classifier (M1), an early-fusion multimodal transformer without retrieval (M2), and a text-only retrieval-augmented clinical language model (M3). The proposed method reaches a detection AUROC of 0.985 and AUPRC of 0.980 with an accuracy of 95.8% and an F1 of 95.2%. For the finer-grained tasks, subtype AUROC/AUPRC is 0.952/0.935 and grade AUROC/AUPRC is 0.934/0.915, with F1 values of 89.6% and 87.5%. Agreement-weighted dual-store retrieval improves Recall@5 and nDCG@5 in both stores and yields measurable accuracy gains attributable to retrieved evidence. Confidence gating lowers the error under abstention to 2.9% and improves calibration (ECE = 0.024; Brier = 0.078). These results indicate that retrieval-grounded multimodal fusion supports accurate, interpretable, and better-calibrated brain tumor evaluation.