Scientific Journal of Astana IT University

EFFICIENT MULTI-ROBOT SLAM IN A LABYRINTH ENVIRONMENT: A CENTRALIZED APPROACH USING ALGEBRAIC CONNECTIVITY AUGMENTATION

2026-03-31T14:56:47+05:00

The visual system is a central component of mobile robot navigation, as it provides essential information about the surrounding environment for motion planning and coordination. However, navigation requires the robot to interpret sensory data through localization and mapping processes, such as Simultaneous Localization and Mapping (SLAM) frameworks. Mapping such environments with a single robot remains a significant challenge due to limited coverage and the substantial time required to map large areas. Multi-robot SLAM offers a promising alternative by distributing the exploration task across several robots, yet practical implementations still struggle with issues such as inconsistent map merging and the overall computational complexity of generating a unified global map. The objective of this study is to implement and evaluate a multi-robot SLAM system in which robots operate independently, and maps are merged on a centralized server in a real labyrinth setting to determine whether collaborative mapping can be practically achieved using mobile turtlebot robots. The system integrates frontier-driven exploration with a two-stage inter-robot loop closure detection process, enabling robots to identify and validate overlapping regions during mapping. Global pose graph optimization is performed on a centralized server that aggregates data from all robots, while pose graph sparsification based on effective resistances reduces the amount of data that must be exchanged. Experiments were first conducted in simulation and later in a physical labyrinth constructed for testing. The results demonstrate that two robots were able to autonomously explore distinct regions, detect inter-robot loop closures, generate sparsified pose graphs, and merge their individual maps into a coherent global representation. The study provides an end-to-end implementation and establishes a foundation for scaling multi-robot SLAM to more complex scenarios.

MODEL OF AN INTELLIGENT GROUP PURCHASE SYSTEM, TAKING INTO ACCOUNT MULTIPLE USER PARTICIPATION AND DELIVERY RESTRICTIONS

2026-06-05T09:42:20+05:00

This study addresses the joint formation of group purchases and allocation of orders to capacity-constrained delivery slots when users may participate in several groups. A mixed-integer optimization model is developed to represent order-slot assignment, group activation, delivery time windows, courier capacity, seller service requirements, and limits on concurrent user participation. The objective minimizes total delivery cost and penalties for unmet seller service levels while preserving feasibility of hard logistical constraints. The model was evaluated on a reproducible synthetic instance comprising 200 users, 40 purchase groups, 300 orders, 10 sellers, and 15 courier slots, each with a capacity of 50 units. Its performance was compared with a baseline assignment procedure that does not jointly enforce the stated restrictions. The proposed model increased average slot utilization from 69% to 85%, reduced average delivery cost from 12.1 to 9.7 conventional units, raised the share of orders delivered within their time windows from 78% to 92%, reduced assignment conflicts from 31 to 11, and shortened average delivery time from 96 to 70 minutes. A two-proportion test confirmed the increase in timely deliveries (z = 4.802, p < 0.001); 95% Wilson intervals were 72.97-82.32% and 88.37-94.57%. An exact conditional test showed a reduction in conflicts (p = 0.0029). Optimality follows from the finite binary feasible set and the branch-and-bound certificate: when the incumbent objective equals the lower bound, no feasible assignment can improve the reported solution. These findings support optimization as a practical basis for robust, reliable, resource-efficient group-purchase delivery planning under competing operational constraints.

MULTILINGUAL NEURAL MODELS FOR TRANSLATING CHAGATAI HISTORICAL TEXTS

2026-05-26T19:22:34+05:00

Recent advances in artificial intelligence have significantly improved automatic translation, yet historical and extremely low-resource languages remain challenging due to limited training data, complex morphology, and orthographic variation. This study investigates the effectiveness of modern neural translation approaches for translating Chagatai historical texts. The written heritage of Central Asia contains many historical documents composed in the Chagatai language, which served as a major literary and administrative language across the region from the Timurid period until the early twentieth century. Although many manuscripts have been digitized, the majority remain inaccessible to broader scholarly and public audiences because reliable translations into modern languages are scarce.

In this study we constructed an experimental parallel dataset by web scraping and custom extraction scripts. Dataset contains 10712 aligned entries in Chagatai paired with English and Kazakh translations. Preprocessing steps includes Unicode normalization, script standardization, and multilingual alignment. After that we use fine-tuning for two different translation architectures: a multilingual NLLB-200 neural machine translation model trained within the No Language Left Behind framework and a TranslateGemma generative large language model instruction-tuned for translation tasks. For the NLLB-200, the tokenizer was extended with a chg_arab token, and training used a Seq2Seq framework with AdamW optimization, mixed precision, and early stopping based on BLEU scores. For TranslateGemma, parameter-efficient fine-tuning with DoRA adapted the model’s key projection and embedding layers to Chagatai orthography, leveraging typological similarity with Uzbek language, while regularization and precision optimizations ensured stable convergence on the limited dataset.

The results show that the NLLB-200 multilingual model consistently produces more reliable translations, achieving higher scores in most evaluation metrics and demonstrating stronger semantic fidelity, particularly when translating into the related Turkic language Kazakh. In contrast, the generative model often produces fluent but less faithful translations and exhibits a higher rate of hallucinated content. For example, XCOMET showed that NLLB-200 consistently outperformed TranslateGemma, achieving 74.97 for English and 74.31 for Kazakh, compared to Gemma’s 73.80 and 69.75, respectively. These findings suggest that source-anchored multilingual translation models, such as NLLB-200, are better suited for translating extremely low-resource historical languages like Chagatai.

AN EFFECTIVE METHOD FOR ANALYZING HUMAN MOVEMENT IN REAL TIME BASED ON YOLO-ROI

2026-07-03T09:20:59+05:00

Real-time human movement analysis has become an essential component of modern computer vision applications, including sports performance assessment, rehabilitation monitoring, intelligent surveillance, and human–computer interaction. Although deep learning–based pose estimation methods provide accurate human body landmark detection, executing person detection on every video frame introduces considerable computational overhead, limiting their applicability on resource-constrained CPU-based systems.

This paper proposes an efficient YOLO–ROI framework for real-time human movement analysis that combines scheduled YOLO person detection, region-of-interest (ROI) generation, MediaPipe Pose estimation, and lightweight landmark smoothing. Instead of performing object detection on every frame, the proposed framework periodically executes the YOLO detector while reusing the most recently detected ROI between detection cycles. Consequently, MediaPipe Pose continuously estimates body landmarks within the ROI, significantly reducing redundant object detection operations and improving computational efficiency.

The proposed framework was evaluated using a self-collected dataset consisting of 200 videos recorded from 10 participants performing five physical exercises under controlled indoor conditions. Computational performance was analyzed under different ROI sizes, detection intervals, and MediaPipe Pose complexity levels using CPU-based execution.

Experimental results demonstrate that the proposed optimization strategy achieves a maximum processing speed of 61.32 FPS with an average processing latency of 16.31 ms, enabling stable real-time operation without GPU acceleration. The obtained results further show that scheduled YOLO detection, ROI reuse, and lightweight landmark smoothing effectively reduce computational workload while preserving continuous pose estimation throughout video processing.

The proposed framework provides a simple, computationally efficient, and easily deployable solution for real-time human movement analysis and can be readily integrated into sports analytics, rehabilitation systems, intelligent surveillance applications, and edge-based computer vision platforms.

RELIABILITY-AWARE ARTIFICIAL INTELLIGENCE FOR MULTILABEL ELECTROCARDIOGRAM DIAGNOSIS

2026-09-15T14:27:50+05:00

Artificial intelligence can classify electrocardiograms, but high discrimination alone does not show whether probabilities, decisions, and explanations remain reliable under technical signal failures. This study evaluates an integrated reliability-aware framework for multi-label electrocardiogram classification on PTB-XL with five diagnostic superclasses: normal, myocardial infarction, ST and T wave change, conduction disturbance, and hypertrophy. Patient-disjoint folds 1–8, 9, and 10 were used for training, validation, and held-out testing, respectively. XGBoost and three one-dimensional deep-learning architectures were compared, and ResNet1D was selected by validation macro area under the receiver operating characteristic curve. On the held-out test fold, ResNet1D achieved a macro area under the receiver operating characteristic curve of 0.9110, macro area under the precision-recall curve of 0.7746, and macro F1 score of 0.7051. Isotonic calibration, fitted on validation predictions only, produced a held-out macro Brier score of 0.08523 and adaptive calibration error of 0.00053. Selective prediction showed a coverage-risk trade-off: higher coverage reduced abstention but increased accepted-case risk and expected harm. Robustness tests covered Gaussian noise, baseline wander, amplitude scaling, lead dropout, and partial waveform failures; grouped limb- and precordial-lead dropout caused the largest discrimination losses. A complete-missing-lead input gate rejected zeroed or flat leads and reduced expected harm under complete lead loss, but did not detect the tested partial waveform failures. Attribution stability was high under Gaussian noise and lower under lead dropout. External analysis on Georgia and Chapman-Shaoxing data was limited to a binary abnormal-versus-normal sanity check and showed source-dependent calibration transfer. The contribution is a reproducible reliability evaluation framework, not a new architecture, full external validation, or a clinical deployment claim.

LYAPUNOV EXPONENT OF EEG SIGNALS AND BRAIN NETWORKS DURING EMOTION PERCEPTION

2026-08-14T14:49:46+05:00

Electroencephalography (EEG) provides a non-invasive means to investigate the dynamic behavior of the brain through electrical activity. Traditional linear analysis methods tend to be insufficient in capturing the complexity of neural signals. This study applies nonlinear dynamic analysis - the Largest Lyapunov Exponent (LLE) - to EEG data recorded during an emotional conflict task in children and young adults. The aim of the study was to define age and sex differences during emotional conflict task performance in Lyapunov exponent of EEG signal. EEG were recorded as participants performed task with viewing facial expressions and hearing auditory words corresponding to four emotions: angry, sad, happy and fearful. EEG preprocessing included filtering, artifact removal by using Independent Component Analysis (ICA). Electrodes were grouped into functional brain networks: anterior, posterior, executive, z-network and left and right hemispheres. LLE was computed from each epoch for broadband to quantify the underlying chaotic dynamics of brain activity, and separately for alpha and theta bands as emotional features. Repeated-measures ANOVA revealed significant differences across brain networks, and interactions with age, sex, and emotions. Machine learning classification found that sex-stratified emotion decoding revealed band-dependent asymmetries: in broadband features, male participants showed higher four-way emotion decoding accuracy (~30.1%) than female participants (~23.5%), while the pattern reversed in the alpha band, where females outperformed males (~27.6% vs. ~19.8%). These interactions suggest that sex differences in LLE-based emotion separability are spectrally specific.

SPATIOTEMPORAL GRAPH PHYSICS-INFORMED NEURAL NETWORKS FOR DATA CENTER THERMAL MODELING

2026-08-14T14:49:26+05:00

Accurate thermal modeling is the precondition for improving energy efficiency and operational reliability of data centers (DCs). However, airflow circulation and workloads cause the temperature field of DCs to exhibit complex spatiotemporal variations. Conventional data-driven models formulate temperature prediction as a multivariate time-series task, without explicitly accounting for spatial heat interactions. In comparison, physics-based methods require detailed geometric information and high computational resources. To address these limitations, we propose a spatiotemporal graph physics-informed neural network (SGPINN) framework for multi-node and multi-step DC temperature prediction. Inspired by resistance-capacitance thermal modeling, SGPINN represents the thermal field as a graph, where graph aggregation captures spatial thermal interactions, and a physical residual combines temperature prediction with graph-guided thermal propagation. The temporal temperature derivative is obtained through automatic differentiation, while learnable equivalent thermal capacitance and heat-flow terms are introduced to characterize heterogeneous node dynamics without requiring directly measured physical parameters. A learnable multi-task strategy is adopted to adaptively balance the data loss and physics loss. The model is evaluated using both a real-world DC dataset and a simulated DC dataset and compared with four representative baseline models. SGPINN achieves the best overall performance on both datasets, with average MAEs of 0.1097 °C and 0.3043 °C, respectively. In the real-world case, it reduces average MAE and RMSE by 3.57% and 3.58% compared with APINN, while ablation results verify the contributions of graph modeling, physical constraints, and adaptive loss weighting. These findings suggest that SGPINN improves predictive accuracy while providing a physics-guided framework for DC thermal modeling.

Ensemble-Based EMG-IMU Gesture Recognition for Real-Time Prosthetic Hand Control

2026-06-25T17:02:50+05:00

Reliable control of upper-limb prostheses remains challenging because muscle signals can change between sessions and wearable EMG sensors may suffer from poor electrode contact. This study presents a real-time prosthetic hand control platform based on an OYMotion gForce Pro armband and a five-finger Dynamixel-actuated prosthetic hand. The system is designed to work without an external dataset, using only a short user-specific recording session and brief calibration before real-time use — specifically, five main training repetitions plus four supplementary calibration repetitions per gesture — intentionally minimising the calibration burden to enable practical single-session deployment.

Five hand gestures—power grasp, open hand, OK gesture, index pointing, and thumbs-up—were classified using several machine-learning and ensemble-based methods. All models were evaluated under the same live data stream to ensure a fair comparison. Six recording sessions collected across five able-bodied participants on different recording days were included in the final statistical analysis. For every session, the main and supplementary training samples were combined to train session-specific models before live evaluation. Across six independent recording sessions from one able-bodied participant, Linear Discriminant Analysis achieved the best average live accuracy of 85.9%, followed closely by majority voting and adaptive ensemble methods. In contrast, the LSTM model showed lower performance, reaching 70.2% accuracy.

The results indicate that, for low-data real-time prosthetic control, ensemble-based methods preserved near-best average accuracy while reducing reliance on any single classifier assumption, making them a more robust deployment choice than individual classifiers. Statistical analysis confirmed that session-to-session variability was the dominant factor in system performance (Kendall's W = 0.249) rather than classifier choice (W = 0.043), suggesting that practical reliability depends more on robust calibration and signal monitoring than on model selection.

TASK-AWARE EVALUATION OF JPEG RECOMPRESSION AND RGB BIT-DEPTH REDUCTION FOR OBJECT DETECTION

2026-08-21T09:58:32+05:00

How does image storage affect object detection? Traditional image quality metrics typically focus on human vision, but computer vision models process images differently. To explore this gap, we evaluated the Microsoft COCO val2017 dataset under two distinct degradation methods: frequency-domain JPEG recompression (quality levels 94 to 25) and amplitude-domain uniform RGB bit-depth reduction (8 bits to 1 bit) stored in 24-bit PNG containers. We processed the degraded images using three distinct architectures: two convolutional networks (YOLOv8n, YOLOv8m) and a Vision Transformer (RT-DETR-L). The results demonstrate that JPEG recompression provides a highly efficient tradeoff; quality level 88 halves the dataset storage size (2.02× compression, reducing to 384 MB) with less than a 1.5% relative drop in mAP50 across all models. Conversely, storing uniformly quantized RGB data in lossless PNG containers proved highly inefficient. While 7-bit quantization preserved baseline accuracy, it artificially inflated the physical file size to 1912 MB. Furthermore, at 3 bits per channel, average mAP50 dropped by over 12%, and all networks effectively failed at the 2-bit and 1-bit thresholds. Crucially, scale-aware metrics revealed that small objects are disproportionately vulnerable to compression artifacts across both CNN and Transformer paradigms, suffering up to a 50% relative accuracy loss. Ultimately, we conclude that standard JPEG compression is significantly more robust for object detection than storing uniformly quantized RGB data in unoptimized containers. These empirical findings highlight that conventional human-centric quality metrics are insufficient for predicting downstream neural network performance. Future research should investigate whether these non-linear degradation trends persist across specialized domain-shifted datasets and modern learned compression algorithms.

DEEP REINFORCEMENT LEARNING CONTROL OF A TENSION-CONSTRAINED TENDON-DRIVEN UNDERACTUATED ROBOTIC FINGER: A COMPARATIVE STUDY OF DDPG AND SAC

2026-09-20T13:13:28+05:00

Tendon-driven robotic fingers are underactuated and strongly nonlinear, and their unidirectional, friction-attenuated tendon tension imposes actuation constraints that any practical controller must respect. A previous study established that a deterministic learning controller (DDPG) can track a grasp trajectory for such a finger and compared it against model-based schemes. The present work extends that line of investigation by asking a narrower but unresolved question: for this same single-actuator, three-link tendon-driven architecture, how does a deterministic actor-critic policy (DDPG) compare with a stochastic, maximum-entropy policy (SAC) once a tension-limit penalty is built directly into the learning objective? Using the previously validated physics-based simulation model, both agents are trained under identical conditions with a reward that jointly penalizes tracking error and tension-constraint violation, and are evaluated across random, sinusoidal, and step disturbances using steady-state error, RMS, ISE, and IAE. The results show a consistent and interpretable trade-off: the stochastic SAC policy attains lower tracking error and better disturbance rejection, especially under time-varying and abrupt loads, at the cost of noisier control and longer training, whereas the deterministic DDPG policy converges faster and produces smoother torque and tension profiles but adapts less well to coupled disturbances. The study isolates the role of the exploration mechanism and the tension penalty in shaping these behaviors and offers practical guidance on selecting between deterministic and stochastic actor-critic methods for physically constrained, underactuated tendon-driven systems.

THE IMPACT OF BUILDING GEOMETRY AND ORIENTATION ON AIRFLOW PATTERNS AND VENTILATION EFFICIENCY IN RESIDENTIAL URBAN ENVIRONMENTS: A NUMERICAL STUDY

2026-07-13T09:14:39+05:00

Air pollution remains a major health concern in urban areas, where building density and layout strongly influence airflow and natural ventilation. While many studies focus on pollutant concentrations and emission sources, fewer examine how building geometry affects the formation of low-velocity zones. This study investigates the influence of building shape and arrangement on ventilation efficiency using two-dimensional numerical modelling. The incompressible Navier–Stokes equations are solved in Python using the projection method on uniform finite-difference grids.

Before analysing the main residential layouts, a test case was performed for two identical domains. In the second domain, the square buildings were rotated by 45°. The results showed that building orientation changes both peak velocity and the extent of low-velocity regions.

Three residential configurations were then examined. Case 1 included L-shaped and rectangular buildings on a 501×501 grid, Case 2 contained only rectangular buildings on the same grid, and Case 3 included diamond-shaped and square buildings on a 301×301 grid. Case 2 showed the smallest low-velocity area (11.1%) and the highest mean air velocity (1.101 m/s), corresponding to a 45% reduction in low-velocity area and a 48% increase in mean velocity compared with Case 1 (20.2%, 0.745 m/s). In Case 3, the diagonal orientation caused local flow acceleration, but the low-velocity fraction remained high at 20.6%.

Overall, the results show that building arrangement has a strong effect on horizontal airflow distribution. They also indicate that peak velocity alone is not sufficient to assess ventilation quality. The proposed two-dimensional approach can therefore be used as an initial screening tool for comparing alternative residential layouts before more detailed three-dimensional simulations are carried out.

EXPLAINABLE MULTISPECTRAL ATTENTION NETWORK FOR SENTINEL-2 AGRICULTURAL LAND CLASSIFICATION

2026-07-13T09:26:59+05:00

Accurate and transparent classification of agricultural land cover from satellite imagery is essential for operational crop monitoring. This study presents MS-AttNet, a compact convolutional network with squeeze-and-excitation channel attention for Sentinel-2 imagery using all thirteen spectral bands. On the EuroSAT benchmark, evaluated with a stratified 70/15/15 split across five random seeds, the model achieves 98.72 ± 0.21% accuracy, a macro F1-score of 0.9868 ± 0.0021, and a mean one-vs-rest AUC of 0.9999 with 1.32 million parameters. A ResNet-50 baseline re-implemented on the identical split achieves 97.90% accuracy with 23.56 million parameters, making the proposed model approximately eighteen times smaller and nearly three times faster at inference. A controlled ablation using five seeds per configuration shows that squeeze-and-excitation modules, convolutional stage depth, and vegetation indices as additional input channels each change accuracy by less than run-to-run variability. This indicates that the benchmark is saturated with respect to architectural choices in this model family; therefore, performance is not attributed specifically to the attention mechanism. Interpretability is assessed using Grad-CAM and SHAP band attributions with quantitative validation. Against a spatially smooth random control, Grad-CAM reduces deletion AUC from 0.857 to 0.738 and increases insertion AUC from 0.864 to 0.916 across 300 test patches. SHAP identifies blue, near-infrared, green, red, and short-wave-infrared bands as dominant contributors, with stable ordering across independently trained models. NDVI and NDWI were used only for spectral-signature analysis and interpretation, not as classifier inputs. The combination of compact size, competitive accuracy, and quantitatively verified explanations supports large-area operational monitoring applications

BILATERAL VISION-LANGUAGE FRAMEWORK WITH CROSS-LATERAL ATTENTION FOR MAMMOGRAPHY

2026-08-25T23:21:51+05:00

Screening mammography is read comparatively: radiologists hold one breast against the other, and several finding categories have no meaning outside that comparison. Most automated pipelines nonetheless encode a single breast at a time. We describe Bilateral MV-CLIP, a contrastive image-text framework in which all four projections of a patient enter one forward pass and the two lateralities exchange information through a Cross-Lateral Attention block, with a learned gate and a placeholder token preserving behaviour when a projection or an entire side is unavailable. Training combines image-text and image-image objectives, graded similarity targets, and a symmetry term tying embedding geometry to recorded asymmetry status. We train on VinDr-Mammo with reports synthesised from structured annotations and evaluate by linear probing of frozen features, cross-modal retrieval and report decoding, against an independently trained single-breast model matched in backbone, optimiser, schedule and epoch budget, with checkpoints selected on a held-out validation partition. On a pre-specified composite endpoint pooling the three comparison-defined categories, fusion did not improve on that baseline: area under the curve 0.781 against 0.792, a difference of -1.1 points with a 95% confidence interval from -9.6 to +7.5. Per-category effects follow the direction the design predicts, but the partition holds 49 positive comparison-defined breasts and no difference survived correction. Retrieval is limited by weak instance-level image-text alignment rather than by report redundancy, with a pool-size-independent ranking accuracy of 0.691. On an external cohort of 11,913 patients labelled by biopsy the representations transfer at reduced discrimination, reaching 0.65 for any abnormality. We therefore report explicit bilateral comparison as plausible but, on this dataset, unconfirmed, and give confidence intervals throughout.

CRITERION-GUIDED AI ARCHITECTURE FOR LEARNING SUPPORT AND PRELIMINARY STUDENT ASSESSMENT

2026-08-13T14:21:44+05:00

This paper presents a pilot evaluation of a criterion-guided human-in-the-loop architecture for an intelligent educational system that supports learning and preliminary assessment of students’ written and oral responses. Instead of assigning a final score directly, the system combines relevant context, formalized criteria, interpretable response features, supporting evidence, uncertainty estimates, and instructor verification. The architecture separates probabilistic components, including retrieval, automatic speech recognition, natural language processing, and a language model, from deterministic assessment and governance mechanisms such as rubric weights, aggregation rules, thresholds, audit logging, and referral rules. Two modes were implemented: a retrieval-augmented generation assistant for formative learning support and an AI examiner for preliminary criterion-based assessment. In the assistant pilot, 336 of 347 queries were processed without recorded technical errors, giving a completion rate of 96.8%; source citations were included in 89.0% of responses, and median response time was 980 ms. Student perceptions of the examiner were evaluated using 52 questionnaires on a five-point Likert scale. Mean ratings were highest for question clarity at 4.81, usefulness of criterion-level feedback at 4.46, and sufficiency of response time at 4.42. Students also reported notable uncertainty about the grade, with a mean rating of 3.79, and concern about speech-recognition errors, with a mean rating of 3.31. Among 45 oral responses, 12 (26.7%) lacked usable word-level timestamps and were referred for review. A worked case demonstrated traceability from response features and evidence to criterion scores, uncertainty handling, aggregation, and instructor verification. A preliminary comparison of 23 paired AI and instructor-released scores was treated as diagnostic because of a severe ceiling effect. The results support technical feasibility and procedural acceptability, but not yet the validity or reliability of automated assessment.

DESIGN AND DEVELOPMENT OF AN INTELLIGENT AGENT-BASED EDUCATIONAL PLATFORM FOR DEVELOPING CRITICAL THINKING IN FUTURE COMPUTER SCIENCE TEACHERS

2026-07-09T15:56:44+05:00

The development of generative artificial intelligence expands opportunities for personalised learning support but also increases the risks of cognitive dependence and weakened critical thinking. This issue is particularly relevant to pre-service computer science teachers, whose future professional practice requires logically grounded reasoning, argument evaluation, and informed decision-making. This study aimed to design, develop, and experimentally evaluate an intelligent multi-agent educational platform for fostering critical thinking among pre-service computer science teachers. The methodology included analysis of scientific approaches to critical thinking development and artificial intelligence agents in teacher education, identification of pedagogical and functional requirements, platform architecture and interaction scenario design, prototype development, and a quasi-experimental study with pre- and post-testing. The developed architecture integrates an intelligent tutoring system for argumentation, role-specific artificial intelligence agents for computer science lesson simulation, an expert retrieval-augmented generation module, and an analytics dashboard. Agent operation is organised through profiling, short- and long-term memory, pedagogically constrained planning, and action modules supporting adaptive task generation, dialogue-based guidance, formative assessment, feedback, reflection, and individual progress monitoring. The platform also enables simulation of pedagogical situations aimed at developing future teachers’ professional readiness to foster school students’ critical thinking. Preliminary experimental results indicate positive changes in critical thinking performance and demonstrate the pedagogical potential of the platform. The next stage will evaluate the computer science lesson simulation mode and its contribution to professional readiness for developing school students’ critical thinking. The practical significance lies in integrating the platform into both subject-specific and pedagogical-methodological components of pre-service computer science teacher education.

Adaptive Learning Methodology for Cybersecurity Culture through Simulations and Gamification

2026-07-07T17:57:56+05:00

People, not technology, are the decisive factor in most information security incidents: phishing and other social-engineering attacks exploit users, not systems. Traditional awareness training — a periodic lecture followed by a test — rarely changes behaviour durably and serves poorly users who differ in age, role and competence, a problem especially acute in Kazakhstan and Central Asia, where digital services are mobile-first and Kazakh-language security education is almost absent. This article develops an adaptive-learning methodology for building a cybersecurity culture across user levels, integrating diagnostic risk profiling, an adaptive trajectory, safety-constrained interactive attack simulations and gamification; a PRISMA-guided bibliometric mapping of 129 works in Web of Science and Scopus grounds the integration gap it addresses. The methodology is formalised as a technical specification and implemented as a deployed, trilingual (Kazakh, English, Russian) web-and-mobile platform whose behaviourally scored simulation logs feed a per-learner risk profile that drives trajectory adaptation. We describe the platform’s core algorithms — action scoring, risk-profile update and weak-spot course recommendation — and report their performance: on the live deployment they execute server-side in under three milliseconds, while authenticated API endpoints respond in roughly 60–85 ms end-to-end. End-to-end operation is verified on a demonstration cohort of 61 learner accounts, 124 course enrolments (34 completed) and 264 graded quiz attempts at a mean score of 76.7%. These results establish feasibility rather than training efficacy, which a randomised controlled trial will measure. The contribution is a reproducible, culture-oriented, Kazakh-inclusive methodology, its technical specification, and a working platform that turns awareness into measurable, adaptive behavioural training.

TABLET-BASED AAC IN A KAZAKHSTANI SPECIAL SCHOOL: IMPLEMENTATION, PRACTITIONER WORKLOAD AND INDIVIDUAL RESPONSE

2026-08-17T16:38:36+05:00

Non-verbal children in Kazakhstani special schools communicate mainly through paper picture cards and gestures. This study followed twelve children aged 6 to 10 with autism spectrum disorder or severe developmental language impairment at Special School No. 3 in Karaganda through the 2024-2025 academic year. Six worked with Proloquo2Go or LetMe Talk on tablets; six continued with cards and gestures. Allocation was not randomised. Three measures were taken: video-coded spontaneous initiation at three points, a symbol test covering taught and untaught items, and the Vineland-3 Communication Domain, completed separately by teachers and parents.

The tablet group moved further on every measure and none of the differences reached significance, which six children per condition cannot deliver for effects of this size. Endpoint initiation averaged 7.4 per fifteen-minute window against 5.1. On untaught symbols the groups scored 42 and 26 per cent, the latter indistinguishable from guessing. Vineland change was 5.2 by teacher rating and 7.8 by parent rating in the tablet group, against 2.9 and 3.7 in the card group. Neither application speaks Kazakh, so every session ran with the device in Russian and the practitioner voicing the child's home language alongside it. This has nothing to do with which method won and everything to do with what deployment here actually involves. The six tablet children also ended the year further from one another than the group averages were from each other. Configuration cost about sixty practitioner hours, mostly in the first two months. For a school weighing this up, the useful question is which children respond, and these data cannot answer it.

EVALUATING A MOBILE APPLICATION FOR DEVELOPING STUDENTS’ ENVIRONMENTAL COMPETENCIES: A QUASI-EXPERIMENTAL STUDY

2026-08-31T11:35:56+05:00

The integration of digital technologies into environmental education creates opportunities to link theoretical knowledge with practical activities and environmentally responsible behavior. The aim of this study was to develop and evaluate the Green IQ mobile app as a supplementary educational tool for fostering environmental competencies among students and to determine whether its integration into traditional environmental education is associated with an increase in students’ environmental competencies. The study involved a quantitative quasi-experimental pre-test/post-test analysis in a control group. Sixty first-year students participated in the study. Participants were divided into an experimental group (n = 29) and a control group (n = 31). Both groups received comparable environmental education through lectures, training sessions, and practical assignments, while the experimental group additionally used the Green IQ mobile app. The Eco Index was assessed using a 10-item instrument developed by the author on a five-point Likert scale, covering environmental awareness, attitudes toward the environment, and environmentally responsible behavior. The results of the study showed that the mean Eco Index for the entire sample increased from 2.53 (SD = 0.769) on the pre-test to 4.03 (SD = 0.727) on the post-test. Both groups demonstrated significant within-group improvement: in the experimental group, it increased from 2.55 to 4.43 (mean change = 1.88), t (28) = 13.10, p < 0.001, Cohen’s d = 2.42, while in the control group it increased from 2.52 to 3.65 (mean change = 1.13), t(30) = 5.35, p < 0.001, Cohen’s d = 0.960. A mixed-effects analysis of variance revealed a significant time-by-group interaction, F(1,58) = 8.39, p = 0.005, indicating greater improvement in the experimental group.

Hardware-Bound Split-Key Framework for Software Protection Against Reverse Engineering

2026-08-25T16:57:16+05:00

Software piracy and reverse engineering cause significant damage to intellectual property: the commercial value of unlicensed software installed worldwide has been estimated at over USD 46 billion. Traditional protection methods rely on static encryption keys or on obfuscation, and they are not resistant to modern analysis techniques such as differential computation analysis (DCA). This study addresses the single point of failure in software protection systems by developing a new key management architecture. The paper introduces a distributed split-key software-hardware framework. A deterministic decryption key is divided into three parts: a local key derived from the unique properties of the machine's hardware (hardware fingerprint), a server key issued after authenticated online activation, and a key sealed in a Trusted Platform Module (TPM) and bound to the platform's Platform Configuration Registers (PCRs). The final key is assembled exclusively in RAM during program execution through an XOR operation and is erased immediately after use. A C#/.NET 8 reference implementation was evaluated on a single Windows platform with a discrete TPM 2.0 module. Changing any of the five hardware identifiers or the PCR state prevented decryption, the assembled key remained in memory for a median of 15.6 µs, key assembly took 0.3 µs, and the one-time TPM sealing took about 67 ms. The runtime overhead stayed below 5% for typical protected functions. The analysis also identifies the remaining limitations of the design: code harvesting through legitimate decryption, binary patching that does not affect PCR 0 and PCR 7, clock rollback during the offline grace period, and loss of access after legitimate firmware or Secure Boot updates. The framework binds execution to authorized hardware and raises the cost of reverse engineering.

LIMITS OF ENTROPY-BASED RANSOMWARE DETECTION: SEPARATING ENCRYPTED FROM COMPRESSED FRAGMENTS

2026-09-19T11:35:50+05:00

Ransomware detectors and forensic tools often treat a file or disk block as encrypted when its Shannon entropy approaches 8 bits per byte. Compressed formats reach almost the same entropy, which causes false alarms, and partial encryption or text encoding of the ciphertext can lower entropy below common thresholds. This study measures how far byte-level statistics separate encrypted from compressed data and how the answer depends on fragment size. A corpus of 595 files in 17 formats, including Deflate, bzip2 and LZMA compressed archives, documents and images, was built from the Govdocs1 collection and encrypted with AES-256 or ChaCha20; evasion variants used intermittent encryption, Base64 encoding and encryption of the first megabyte. Fragments of 512 bytes, 4 kilobytes and 64 kilobytes were described by entropy, the chi-square statistic and four further byte statistics, and threshold rules, logistic regression and random forest were evaluated with cross-validation grouped by file. Entropy and chi-square ranked fragments almost identically, as expected from their known asymptotic equivalence, so chi-square did not separate the classes better; its advantage was a size-independent threshold. A fixed entropy threshold of 7.9 bits per byte missed every encrypted 512-byte fragment and flagged 73% of unencrypted 64-kilobyte fragments. Deflate-based formats became separable only with large fragments, whereas xz and 7z archives remained close to chance for every method. Base64 encoding removed detection completely, and intermittent encryption was rarely detected in large fragments of uncompressed files. Byte statistics can therefore show that data are not plaintext but cannot confirm that compressed data are encrypted; detectors need size-aware thresholds, decoding of text encodings and format validation.

NEW APPROACH OF BYTE-LEVEL RECONSTRUCTION FOR ENHANCED DIGITAL FORENSIC DATA RECOVERY BASED ON ML

2026-08-25T23:23:56+05:00

The exponential growth of digital data and the increasing sophistication of cyber threats have made traditional data recovery methods, which rely on file system metadata and signature-based carving, inadequate for modern forensic investigations. Solid-state drives with TRIM and wear leveling, file fragmentation and partial corruption frequently leave investigators with incomplete evidence and without any quantitative measure of its reliability. The aim of this study is to develop and validate a machine learning-based ensemble framework for intelligent byte-level data recovery and integrity assessment. The proposed method extracts an 18-dimensional feature vector describing the statistical, structural, entropy and pattern properties of each file segment and combines three complementary models: a Random Forest that predicts individual byte values from contextual windows, a Gradient Boosting model that reconstructs pattern-level structure, and a multilayer perceptron that estimates structural recovery confidence. Their outputs are merged through weighted voting with weights optimized by cross-validation, followed by structural validation and a multi-metric confidence assessment. The framework was evaluated on 10,247 files (PDF, JPEG, PNG, DOCX, MP4) stored on NTFS, ext4 and APFS and subjected to 15 synthetic corruption scenarios, including bit errors, block deletion, header damage, fragmentation, zero padding and pseudo-encrypted regions. The ensemble achieved 89.3% recovery accuracy and 92.1% precision, substantially surpassing Autopsy (62.4%), FTK Imager (58.7%) and EnCase (74.5%), and all improvements were statistically significant (p < 0.001). Accuracy remained stable across file systems (87.1–91.7%), and the confidence scores were well calibrated, with an Expected Calibration Error of 1.48%. The price of this gain is a longer processing time of 45.2 s per file. The results show that combining ensemble byte-level reconstruction with calibrated integrity assessment yields more complete evidence together with quantitative confidence measures suitable for legal proceedings, providing a practical complement to existing forensic tools.

FORENSIC RECOVERY OF VIDEO EVIDENCE FROM DAHUA DHFS4.1 SURVEILLANCE SYSTEMS AFTER FORMATTING

2026-08-27T23:34:21+05:00

Video footage from proprietary surveillance systems is central to criminal investigations, yet its forensic recovery remains poorly supported. Conventional carving fails on Dahua systems because the proprietary DHFS4.1 file system uses frame encapsulation, header-footer validation pairs and embedded checksums. This paper addresses two open questions: how each of the four DHFS4.1 validation levels affects the false discovery rate (FDR) and recovery rate, and how quick formatting and partial bulk zero-fill overwriting differ in their impact on recovery. The technique combines dual-signature frame validation, frame-size and checksum consistency checking, and adaptive time sequencing, extending our earlier pipeline; the principal contribution is the validation-level ablation, the paired comparison against three independent tools, and the partial-overwrite study. All methods were executed on forensic images of eleven Dahua NVR hard drives acquired behind a hardware write blocker, so every comparison is paired within drive. The proposed method achieved a recovery rate (RR) of 88.0%, a temporal accuracy (TA) of 89.0% and a file-level FDR of 4.0%; four-level validation reduced the frame-level FDR 13.5-fold, from 27.0% (single signature) to 2.0%, at the cost of a 3.2-percentage-point drop in RR. The RR advantage over each compared tool was statistically significant (paired-samples t-test, n = 11, p < 0.05) and confirmed by the Wilcoxon signed-rank test. Quick formatting still permitted 76% recovery, whereas bulk zero-fill reduced it to 34% and raised the file-level FDR to 41%; a collision analysis shows that this increase stems from payload fragmentation rather than chance signature matches. Within the tested scope, the methodology performs comparably to commercial tools and, because every decision rule is published, is re-implementable and transparent. The findings offer practical guidance for prioritising evidence collection from Dahua surveillance systems.

RECORD-LEVEL VALIDATION OF OPEN-SOURCE UAV FLIGHT LOG PARSERS ON THE NIST CFREDS CORPUS

2026-09-18T11:18:48+05:00

Examiners read drone flight logs with open-source parsers, and published comparisons of these parsers count the rows a tool writes or note whether it crashed. We tested whether such counts say anything about the evidence. Six parsers (DatCon, DROP, DRDP, Fodogu, pymavlink and GRYPHON) were run three times on each of the 72 files of four NIST CFReDS drone datasets, 1,296 runs with pinned versions and hashed inputs and outputs. Every coordinate a tool wrote was compared with two reference decoders: one accepts a DJI packet only when its checksum verifies, the other reproduces the Mission Planner exports shipped with the ArduPilot logs. On the DJI Phantom 3 files, DatCon and DROP wrote no unsupported position and agreed exactly wherever both reported one; 729,793 reference records (82.4%) are reproduced by both. The count differences earlier read as complementary recovery come from repeated rows and from two undocumented display rules: DROP blanks positions logged with two or fewer satellites, and DatCon writes nothing before a flag set by the first record with non-zero position and height. At the rate used here, 98.3% of DatCon's GPS-receiver rows repeat an earlier value (97.9% at its 30 Hz default). Runs that ended with status zero and no coordinates from a file holding them made up 21 of 126 runs on formats a tool claims to support and 61 of 120 on other formats. On the ArduPilot logs, 545 of 927 positions that pass a plain latitude and longitude range test were logged without a three-dimensional fix. All 431 completed run triples were bit-identical. The procedure is stated as a formal model, so that each reported quantity has a definition.

VOTINGONTOSEC: A SECURITY ONTOLOGY FOR RISK MODELING IN ELECTRONIC VOTING SYSTEMS

2026-08-31T11:01:59+05:00

Blockchain voting needs a security model for shared infrastructure, cryptographic tools, and election rules. General ontologies do not jointly cover its threats and standards-based risk assessment. This paper presents VotingOntoSec, an OWL ontology aligned with ISO/IEC 27005. It represents assets, threats, vulnerabilities, risks, and countermeasures, including Sybil and smart-contract attacks and cryptographic protections. Development in Protégé 5.6.1 was followed by graph-based structural tests, HermiT 1.4.3 reasoning, competency-question assessment, comparison with five representative ontologies, and a Smart City voting scenario. The ontology contains 118 classes, 16 object properties, and 9 data properties. Its class hierarchy has no cycles, and HermiT found all named classes satisfiable. A five-asset example illustrates traceability from a vulnerable asset through an exploiting threat to a suitable control. Expert-assigned likelihood and criticality estimates yield reproducible inherent and residual risk scores. In the illustrative calculation, the criticality-weighted system score decreases from 0.81 to 0.32 after applying modeled controls; rankings remain unchanged as the likelihood weight varies from 0.3 to 0.7. Among the compared models, VotingOntoSec uniquely combines voting-specific blockchain threats, cryptographic countermeasures, and ISO/IEC 27005-aligned risk analysis. The case includes voter data, the ledger, consensus nodes, a communication channel, and a voting device. Results help rank needed controls and show how risk may change as inputs or weights change. The numerical example is illustrative rather than an empirical performance evaluation. The ontology and scripts are available for independent verification and reuse. VotingOntoSec provides a new practical semantic basis for structured security assessment of decentralized voting systems at scale.

ADVANTAGES AND DISADVANTAGES OF THE DIFFUSION PROPERTIES OF LIGHTWEIGHT BLOCK CIPHERS BASED ON ROUND-BY-ROUND AVALANCHE ANALYSIS

2026-05-05T16:30:31+05:00

This paper investigates the diffusion properties of lightweight encryption algorithms designed for deployment in resource-constrained systems. Modern schemes intended for environments with strict limitations on computational complexity and energy consumption are evaluated. Particular attention is paid to the analysis of isolated round functions ‒ particularly their behavior prior to the application of round keys. This approach enables the assessment of internal structural properties independently of the key scheduling procedures and identifies basic diffusion characteristics determined exclusively by the employed bitwise transformations. Specifically, the impact of flipping a single bit in the input block on the distribution of changes in the output data is analyzed, facilitating an evaluation of the quality of internal transformations and their resistance to cryptanalytic attacks. For quantitative evaluation, classical measures of avalanche behavior – the Avalanche Effect and the Strict Avalanche Criterion (SAC) – are used to assess how rapidly and uniformly information propagates within a block after the application of round transformations. All algorithms are evaluated using a unified round-by-round experimental procedure under identical conditions. This ensures a consistent and objective comparison of their diffusion characteristics across successive encryption rounds. The results reveal distinct advantages and limitations in diffusion speed, stabilization behavior, and uniformity among the evaluated lightweight cipher constructions. The comparison shows that schemes based on cyclic rotations and bitwise operations exhibit high diffusion rates with minimal computational complexity. This analysis demonstrates the effectiveness of approaches relying on simple bitwise transformations. The obtained results provide valuable practical design insights for the selection or design of lightweight cryptographic primitives for microcontrollers, sensor networks, and other devices with limited computational capabilities.