Problems of Informatization and Control

Adaptive approach to load balancing with QoS guarantees in an Edge-Fog-Cloud environment

2026-05-26T09:49:36+03:00

The article proposes an adaptive approach to load balancing in heterogeneous Edge-Fog-Cloud computing environments ensuring Quality of Service (QoS) guarantees. A mathematical model of a three-tier infrastructure is developed, formalizing the characteristics of computing nodes, task flows, and QoS constraints. A node priority evaluation function is proposed, taking into account computing load, network latency, energy efficiency, and reliability; the weight coefficients are dynamically adapted to the task class. The algorithm makes decisions on task placement across Edge, Fog, and Cloud tiers based on the current system state and task QoS requirements. Simulation results in the iFogSim environment compared to Round Robin, Random, and Min-Min algorithms demonstrate a 36–44% reduction in average execution latency, a 30–37% decrease in energy consumption, and an increase in the QoS compliance ratio to 89–98% under various load scenarios.

Hybrid method for anomaly detection in logical clusters of cyber vulnerabilities using an ensemble of machine learning models

2026-05-26T09:45:58+03:00

The article presents the hybrid anomaly detection method based on logical clustering of vulnerabilities and the application of machine learning techniques. The dataset includes information about software vulnerabilities, including CVSS metrics, severity levels, CWE types, temporal characteristics, and assignment to logical groups formed based on the OWASP Top Ten.
At the first stage, data analysis is performed along with the construction of additional temporal features. Subsequently, logical clustering of vulnerabilities is carried out, allowing the data to be partitioned into semantically homogeneous groups such as Broken Access Control, Injection, Cryptography & Data Protection, Configuration & Design Weaknesses, Authentication and Mishandling Failures, and Software Failures. This approach reduces data heterogeneity and improves the quality of subsequent analysis.
At the second stage, the Isolation Forest anomaly detection method is applied to identify atypical vulnerabilities within each cluster. The obtained results are used as pseudo-labels for training ensemble machine learning models, including Decision Tree, Random Forest, and Gradient Boosting. These models are employed for anomaly classification and for evaluating their predictability based on a feature space that includes CVSS metrics, exploitability score, impact score, and temporal characteristics.
Model performance is evaluated using Accuracy, Precision, Recall, F1-score, and Log Loss metrics. Experimental results show that Gradient Boosting demonstrates the best performance on large datasets, while Random Forest provides more stable results on smaller clusters. Decision Tree offers high interpretability but is prone to overfitting and exhibits weaker generalization capability compared to other models.
The proposed hybrid method improves anomaly detection accuracy through the use of semantically homogeneous clusters and an ensemble of models.
The results of the study can be applied in information security monitoring and incident management systems, as well as in the development of intelligent forecasting systems aimed at detecting anomalous events and modeling the dynamics of vulnerability propagation.

Analysis of modern computer vision methods for monitoring and identifying flying objects: challenges and prospects

2026-05-26T09:43:55+03:00

This review article presents a comprehensive analysis of modern software and algorithmic computer vision methods used for monitoring and identifying flying objects in airspace. The relevance of the study is driven by the rapid development of unmanned aerial vehicles and the growing need for efficient, affordable, and highly accurate surveillance systems capable of operating under uncertainty, limited resources, and complex backgrounds. The classification of the main computer vision tasks is considered, including object detection, classification and identification, localization and tracking, as well as estimation of the spatial position and behavioral characteristics of targets. The evolution of detection methods from traditional approaches to modern deep learning architectures is analyzed, including convolutional neural networks, one-stage and two-stage detectors, visual transformers, and specialized approaches for detecting small-sized objects. Particular attention is paid to the problem of distinguishing birds from drones, multi-object tracking methods, data association algorithms, and the robustness of systems against adversarial attacks. A comparative analysis of the effectiveness of computer vision methods is also carried out.

Concept of information flow analysis technology within OSINT operations in instant messaging systems

2026-05-26T09:40:23+03:00

This document presents the results of an empirical study on the efficiency of hyper-automated information systems in processing large volumes of data during Open Source Intelligence (OSINT) operations. The focus is on the development and experimental validation of a conceptual architecture for a next-generation OSINT technology based on integration: Robotic Process Automation (RPA), machine learning tools, and large language models (LLM, particularly LLAMA 3). The technology is applied to data collection from instant messaging systems. The key stages of the hyper-automation pipeline are described: collection, preprocessing, analysis, and classification of data from Telegram channels. RPA bots were used for automated news extraction, followed by linguistic model processing to generate informative reports.
The experiment demonstrated a significant reduction in data processing time (from 6 hours to 17 minutes) while maintaining high accuracy (98%). The results were compared with manual processing and the potential for the proposed approach in information security and disinformation detection was evaluated. The findings indicate the strong potential of hyper-automation in intelligent information processing and data-driven decision-making.

Approach to accelerate computer realization of modular exponentiation of long numbers

2026-05-26T09:37:50+03:00

The article proposes an accelerating method of base cryptography with open key operation - modular exponentiation while being realized on terminal computer platforms of remote controlled systems for real-worlds objects. Reducing the time of exponentiation can be achieved by accelerating modular multiplying due to use of pre-computations that allow combining additions to a sum of partial products of shifted multiplicand and a modular correction code. The proposed method differs in that the pre-computations are being executed in every cycle of exponentiation and being used for accelerating both modular squaring and multiplying operations.
It has been theoretically showed and experimentally proven that proposed method allows an acceleration of calculating the modular exponent by more than twice.

Analysis of the basic mathematical methods used for description of the motion of an object on a plane and in the space

2026-05-26T09:35:01+03:00

The article provides a comparative analysis of key mathematical tools in linear algebra used for modeling the orientation and motion of objects. The study examines the specific applications of Euler angles, rotation matrices, and quaternions within the context of autonomous vehicle control. The advantages of each of these approaches are discussed. Some attention should be paid to the methodology of teaching the seconcepts to future IT specialists and aviation engineers using modern information technologies.

Organization of countermeasures against attacks on cryptographic keys by analyzing the power-consumption dynamics of IoT terminal platforms

2026-05-26T09:32:26+03:00

The paper addresses the problem of improving the protection of modular exponentiation, which is the basic operation of public-key cryptography in IoT-based remote control systems, against attacks based on power-consumption analysis. It is shown that classical modular exponentiation algorithms are vulnerable to SPA, DPA and CPA because the sequence of executed operations depends on the bits of the secret exponent.
A method of polymorphic modular exponentiation is proposed, in which the exponent bits are processed in variable-length groups using precomputations. The use of different group sets for each computation eliminates the direct relationship between the exponent bits and modular multiplication operations, which complicates power-analysis attacks and prevents statistical reconstruction of the exponent code.
It has been proven that, in addition to improving the protection level, the proposed method provides a 7–15% acceleration of software modular exponentiation, which is important for low-power IoT terminal platforms.

Interface architecture of a modular video analysis platform for autonomous uav control

2026-05-26T09:30:19+03:00

The purpose of this work is to present the interface architecture of a modular software platform for video analysis in autonomous UAV control systems, with emphasis on extensibility and component replaceability.
The platform is organized around three independent modularity levels. The first is the video source contract (CameraSource): regardless of the camera type, each source runs a background capture thread and writes frames to a shared buffer. Three implementations are available – USBCameraSource via OpenCV VideoCapture, PiCameraSource using the Picamera2 library for Raspberry Pi CSI cameras (including infrared models), and VideoFileSource for offline development with recorded footage. Switching between them requires changing a single line in the server configuration. The second level is the frame processing contract (FrameProcessor), built on duck typing rather than formal inheritance. A researcher implements one mandatory method – photo_processing(frame) – which receives the current BGR frame as a NumPy array. Three optional methods extend the interaction: catch_cmd(roi) is called when the operator clicks on the video stream to select a region of interest; customize_processing(id) responds to one of four control panel buttons, enabling in-flight algorithm parameterization; getPreparedVector() closes the autonomous control loop by returning velocity components vx, vy and execution time t. The third level covers flight controller connectivity: the FlightController class wraps MAVSDK and accepts a MAVLink connection string as its only configuration parameter, enabling transparent switching between Pixhawk, SpeedyBee, Matek hardware and the SITL software simulator. Four FrameProcessor implementations – a CSRT tracker with SIFT initialization, SIFT homography-based position holding, YOLO object detection, and a minimal template – were integrated and validated on the platform, each implemented as a single class without touching any other component.
The proposed architecture lowers the barrier for integrating new video processing algorithms to implementing a minimal class interface. A researcher working with the platform needs to focus only on the algorithm itself – camera handling, telemetry logging, and flight command execution are already provided. This makes the platform a practical and reproducible baseline for vision-based UAV autonomy research.

Dynamic routing method in mobile software-defined networks using artificial intelligence

2026-05-26T09:22:28+03:00

The article examines the problem of ensuring effective routing and load balancing in mobile Software-Defined Networks (SDN), which are characterized by high node mobility and frequent topology changes. Existing solutions based on classical static approaches or basic artificial intelligence models (for example, multilayer perceptron) have a significant drawback: they analyze the network state as a snapshot at a fixed moment in time and are unable to capture temporal changes and the evolution of the topology graph. This leads to SDN controller overload and reduced Quality of Service (QoS) due to constant link breaks. To solve this problem, a conceptual architecture for a predictive routing system based on Temporal Graph Neural Networks (T-GNN) is proposed. The suggested approach allows for modeling spatio-temporal dependencies in the network, predicting the future state of connections based on historical data, and establishing routes proactively, which significantly reduces the load on the control plane.

Evaluating the capability of machine learning models to detect unknown cyberattacks at digital substations

2026-05-26T09:20:01+03:00

The article addresses the problem of evaluating the capability of machine learning models to detect unknown cyberattacks in network traffic of digital substations. The relevance of the study is determined by the fact that digital substations are part of critical infrastructure, and their operation depends on secure network interaction between automation devices, monitoring systems and industrial communication protocols. The experimental study is based on the processed IEC 104 part of the SANDI-2024 dataset, which was designed for training and evaluating intrusion detection systems for electrical substations. The paper proposes a leave-one-attack-out evaluation scenario, in which one attack type is completely excluded from the training set and used only during testing. This approach makes it possible to evaluate the ability of models to detect attack types that were not represented in the training data. The study compares supervised machine learning models and anomaly-based approaches. The supervised group includes Logistic Regression, K-Nearest Neighbors, Linear SVM, Gaussian Naive Bayes, Random Forest, Extra Trees and HistGradientBoosting. The anomaly-based group includes Isolation Forest, One-Class SVM and Local Outlier Factor. The experimental results show that in the standard known-attack scenario, the models achieve nearly perfect classification metrics. However, in the leave-one-attack-out scenario, model performance depends significantly on the type of unseen attack and on the decision threshold. Extra Trees achieved the best average result among supervised models, while Local Outlier Factor demonstrated the best balance among anomaly-based models. Additional threshold analysis showed that the detection of the most difficult unseen attack can be significantly improved by adjusting the classification threshold. The obtained results confirm the importance of using unseen attack detection scenarios for a more realistic evaluation of IDS models in digital substations.

Parametric optimization in design of small uav flight control systems

2026-05-26T09:17:48+03:00

This paper is devoted to the issues of flight control systems design for small Unmanned Aerial Vehicle (UAV) in the presence of incomplete measurements and stochastic disturbances. To solve this problem the parametric optimization procedure based on the Nelder-Mead method is proposed. With the help of this procedure it is possible to estimate the desired location of poles of the Luenberger filter (used for unmeasured states restoration) and elements of weight matrices that are needed for optimal deterministic controller design. To convert the constrained optimization procedure to unconstrained one it is proposed to include the penalty function into the performance index that is minimized. The efficiency of the proposed parametric optimization procedure in problems of flight control systems design is proved on the model of small UAV with the help of simulation.

Analysis of inventory management issues in warehouse logistics of manufacturing enterprises based on a systematic approach

2026-05-26T09:15:10+03:00

The article examines inventory management in warehouse logistics for manufacturers using a systematic approach. It analyses the warehouse's economic role, showing it is more than storage; it supports production, improves service, and impacts logistics costs. The article also distinguishes between independent and dependent demand inventories, outlining their roles in managing recurring, safety, transit, and seasonal stock.
The article explains in detail the main functions of inventory management systems: accounting and data collection, planning and forecasting, order fulfilment, operational control, and performance evaluation. It also analyses specific factors affecting inventory levels in warehouse logistics: demand variability, supply time, internal operational capacity, resource constraints, and the quality of information systems.
The study uses mass service systems (MSS) and GPSS World (General Purpose Simulation System World) discrete-event simulation as its methodological foundation. By combining GPSS simulation with the analytical MSS approach, the study identifies shortcomings in inventory management systems. This integration optimises inventory levels and improves warehouse logistics efficiency within manufacturing enterprises.

Efficiency of TLS 1.2 and TLS 1.3 cryptographic protocols in modern web applications: analytical and practical aspects

2026-05-26T09:12:41+03:00

This paper presents an empirical evaluation of TLS 1.2 and TLS 1.3 cryptographic protocols in modern web applications across two dimensions: cryptographic strength and system performance. An analytical review traces the protocol evolution from SSL 3.0 to TLS 1.3 (RFC 8446), documenting the transition from vulnerable CBC modes to AEAD algorithms (AES-256-GCM and ChaCha20-Poly1305) and the 1-RTT handshake with mandatory ECDHE key exchange.
The study compares two practical deployment scenarios of Nginx 1.18 with OpenSSL 3.0.2 on Ubuntu Server 22.04 LTS: a typical server-side configuration (TLS 1.2 + AES-256-GCM with AES-NI acceleration) and a mobile/IoT-oriented configuration (TLS 1.3 + ChaCha20-Poly1305 without AES-NI, emulated via OPENSSL_ia32cap). An additional control configuration (TLS 1.3 + AES-256-GCM without AES-NI) was included to compare cipher-suite behavior within TLS 1.3. Measurements were performed using open-source tools — wrk, Wireshark and openssl s_time.
The TLS 1.3 scenario demonstrated handshake time reduction from 4.8 to 3.0 ms (−37 %), TTFB reduction from 48.3 to 30.1 ms, and throughput increase from 430 to 515 Mbit/s (+20 %). The observed −18 % CPU utilization difference reflects the combined effect of protocol optimization, cipher choice and software-only execution — representative of the asymmetry between server and client-side TLS deployments. A practical 13-item secure TLS deployment checklist for DevSecOps engineers is proposed, aligned with NIST SP 800-52r2 and the Mozilla Server Side TLS Intermediate profile. The paper outlines the quantum-resistant transition path via NIST FIPS 203 (ML-KEM) standardization and IETF hybrid key exchange drafts (X25519MLKEM768, SecP256r1MLKEM768).

A method for determining the collective emotional state based on an emotional network and a FER-model

2026-05-26T09:10:40+03:00

This paper proposes a method for determining the emotional state of a group of people based on the construction of an emotional network and dynamically weighted graphs. To assess the emotional state, the Aggregated Emotional Profile (AEP) and Dominant Group Emotion (DGE) metrics are employed, enabling the formation of a generalized emotional profile of the group. The nodes of the graph correspond to individual persons, while the edge weights reflect the degree of emotional similarity or mutual influence between them. Positive, negative, and neutral emotional clusters are formed based on emotional vectors. The proposed approach enables the analysis of collective emotional dynamics, the identification of conflict nodes, and the prediction of behavioral scenarios within groups of people in the future.

Generalized algorithm for assessing pilot preparedness for special flight situations

2026-05-26T09:07:32+03:00

Currently, significant attention is paid to the training of flight personnel for emergency flight situations. During training on full-flight simulators, various scenarios involving aircraft system failures are practiced. However, it is practically impossible to simulate the simultaneous occurrence of multiple avionics failures. Therefore, in addition to rehearsing specific scenarios, it is necessary to conduct anti-stress training. This training involves simulating the simultaneous impact of multiple adverse factors in a real flight (factorial overlaps) by introducing three concurrent failures on a full-flight simulator. This article is devoted to optimizing the application of methods developed at the Department of Avionics and Control Systems for preparing aircraft crews for flights under emergency conditions. Schemes have been developed for creating functional evaluation cards for assessing flight personnel performance by instructors.

Overview of effective methods and tools for building information systems for monitoring messages in the mass media

2026-05-26T09:05:26+03:00

The article analyzes the main methods and tools for building information systems for monitoring mass media. The main features of the system are considered, capable of processing large data sets and analyzing the frequency and context of news. The author proposes a methodology for collecting and analyzing information from diverse sources, and also evaluates them using natural language processing methods. The algorithms for recognizing named entities, modeling and automatic message identification using time series, Markov processes and criteria for their evaluation are analyzed. The research results allow to increase the level of accuracy and identification of information systems through the operational detection of risks. The practical significance of the article lies in using the proposed methodology as an effective tool for managing information systems, media space, and digital communications.

Algorithm for load balancing in distributed control systems for large-scale data streams in real time

2026-05-26T09:01:08+03:00

The article analyses algorithmic methods of load balancing in distributed control systems that process large volumes of data streams under real-time constraints. As flow rates and architectural complexity increase, the operational goal is to achieve throughput and limited latency while ensuring even utilization of heterogeneous computing resources. Previous studies show that in real-time environments, misallocation of work leads to queues, long wait times, and delays in message processing. The main causes of imbalance are peak loads, rapid changes in flow intensity, concentration of events around “hot” buttons, and hardware heterogeneity in nodes. State-of-the-art load balancing strategies are systematized, including static, dynamic, and adaptive algorithms, as well as feedback-based mechanisms. The limitations of static policies in streaming settings are emphasized, and the foundations of algorithms capable of quickly responding to system state changes are established. A conceptual adaptive approach to load balancing based on online node performance indicators, including task queue occupancy, CPU load, and task execution time, is proposed. The performance of the balancing strategies is evaluated using throughput, p95 and p99 delay quantiles, node-level load uniformity, and the proportion of lost messages. The results show that adaptive load balancing improves the reliability of distributed systems by reducing the risk of congestion and stabilizing queuing delay during heavy flow processing. Further research should focus on workload prediction, context-sensitive threshold adaptation, and environment-sensitive reallocation policy design for streaming manufacturing platforms.

Evolution of AI agent architectures and the challenges of optimizing computational resources in online learning services

2026-05-26T08:51:13+03:00

The article provides a systematic review of the evolution of AI agent architectures in distributed intelligent systems, focusing on resource intensity, scalability, and ensuring QoS/SLA in online learning services. It analyzes the transition from monolithic implementations to multi-loop compositional architectures, distinguishing inference, orchestration, memory, and observability. The study generalizes classes of resource optimization methods at the model level, inference serving, autoscaling, flow management, and agent logic. A research gap is identified: the lack of an integrated method that combines model-level optimization, inference serving, and autoscaling with the trajectory of an agent session as the primary unit of resource accounting and control—that is, aligning the token-and-tool profile of a session with admission, prioritization, and scaling policies in a multi-tenant environment. Preconditions are formulated for developing a method and a tool for optimizing the computational resources of AI agents in online learning services.

Metrological support of information-measurement systems for monitoring the technical condition of electrical machines

2026-04-27T12:24:03+03:00

The article addresses the issues of metrological support of information-measurement systems used for monitoring the technical condition of electrical machines. The main sources of systematic and random errors in measurement channels of currents, voltages, and winding temperatures are analyzed, as these errors directly affect the reliability of diagnostic decisions. Mathematical models of measurement channels are developed, taking into account dynamic operating modes of electrical machines and thermal instability of measurement system components. A comprehensive approach to evaluating the total measurement error of the information-measurement system and its influence on the accuracy of technical condition assessment is proposed. The obtained results can be applied to improve diagnostic reliability, measurement accuracy, and operational efficiency of electrical machines in industrial electric drive systems.

Semantic scene graph as a foundation for marker-free AR object spatial anchoring

2026-04-27T12:03:19+03:00

This paper proposes the SG-Anchor algorithm for marker-free spatial anchoring of augmented reality objects on the Meta Quest 3 platform. The method employs a semantic scene graph — a directed weighted graph G=(V,E,W,F,P) where vertices represent semantic objects of the physical environment detected by YOLOv8, and edges encode their spatial-semantic relations. Vertex attributes are formed by concatenating semantic embeddings, geometric bounding box parameters, and statistical stability features. A Graph Attention Network is used for subgraph matching during anchor resolution.

Theoretical analysis demonstrates the method advantage over SLAM-based approaches in scene-change robustness (SRI approx 88-92%) at an expected positioning error of TE approx 2.1 cm.

А method for reducing routing overhead in AD-Hoc networks based on SNR level analysis

2026-04-27T11:29:39+03:00

Increasing node density in wireless Ad-Hoc networks leads to excessive routing control overhead, reduced scalability, and higher energy consumption. This paper proposes a method for logical routing density reduction based on a combined analysis of Signal-to-Noise Ratio (SNR) and node degree. A mathematical model is developed to describe the relationship between node density, average degree, and routing overhead. To reduce effective routing participation without altering physical topology, a combined eligibility metric integrating normalized SNR and node degree is introduced. The proposed algorithm performs threshold-based link filtering to construct a sparsified routing graph. Simulation results for networks with 50–200 nodes demonstrate that the method reduces routing control overhead by up to 35% and decreases average logical node degree by 15–30%, while maintaining stable Packet Delivery Ratio. The increase in end-to-end delay remains within acceptable limits. The proposed approach improves routing scalability in dense Ad-Hoc networks while preserving connectivity and Quality of Service.

System for optimizing the schedule of cases taking into account user preferences

2026-04-27T10:09:52+03:00

The work is devoted to the development of methods for automated formation of personal and work schedules in which tasks have different priorities, time constraints, and types of execution. The study reviewed modern calendar services and task management systems, identified their key limitations, and proposed a new approach to building an optimized schedule. The developed method combines greedy algorithms for forming the initial schedule, local improvement methods, and adaptive task redistribution, which ensures minimization of overlaps, efficient use of user time, and stability of the schedule when it changes.

The proposed approach allows for individual priorities, fixed events, task importance, possible time windows, and user behavior patterns to be taken into account. As a result, the schedule formation is closer to real conditions, and the system itself is capable of increasing planning efficiency, reducing time losses, and ensuring high accuracy of adaptation in case of dynamic changes in the schedule.

Technology for assessing the quality of program systems components based on code complexity metrics

2026-04-27T08:44:38+03:00

The article investigates methods and tools for assessing the internal quality of software using code complexity metrics according to M. Halstead and T. McCabe. A general analysis of the maintainability quality characteristic and its subcharacteristics from the SQuaRE ISO/IEC 25010 standard is performed, which can be effectively used to assess the maintainability of program systems. The main attention is paid to the Halstead and McCabe metrics, since they allow quantitatively assessing the structural complexity of the code of software components. It is shown that these metrics are useful and effective for determining the quality measures of analyzability, modifiability and testability of software modules. An approach to interpreting the results of static code analysis based on the SQuaRE quality model is proposed. The practical part is illustrated by an example of Python code.

Adaptive hybrid transformers for controllable audio synthesis via representation alignment and dynamic modality weighting

2026-04-27T08:35:43+03:00

This article proposes an adaptive hybrid transformer framework for controllable audio (Foley) synthesis that addresses the persistent “control gap” between user-intended perceptual attributes (e.g., pitch and intensity) and the characteristics realized in diffusion-based generative latent spaces. The method integrates three complementary mechanisms: Gated Cross-Attention (GCA) to stabilize multimodal fusion and suppress irrelevant visual tokens, mitigating attention collapse and attention-sink behavior, Dynamic Attention Fusion (DAF) that assigns context-dependent modality weights using normalized Shannon entropy as a differentiable reliability proxy, improving robustness under modality degradation (e.g., visual noise or vague prompts); and improved Representation Alignment (iREPA) that distills structural knowledge from frozen teacher encoders to accelerate convergence while preserving spatial/temporal structure relevant to synchronization. For parameter-efficient controllability, the framework employs LoRA/MoE-LoRA adapters as functional control bases, enabling fine-grained manipulation of acoustic attributes with minimal additional parameters. Quantitative evaluation uses controllability-specific metrics (CSS/COI) and automated validation via AuditEval-ssl, demonstrating strong correlation with expert ratings and improved robustness in combined-noise scenarios.

Review of UAV optical navigation methods and algorithms

2026-04-27T08:28:21+03:00

The article is devoted to a review of navigation methods for unmanned aerial vehicles (UAVs) via the optical channel. The relevance of the topic is due to the fact that optical navigation ensures UAV autonomy in the absence of a GNSS signal, which is a typical problem in modern conditions. The paper examines and compares the main methods and approaches used to implement optical navigation. Modern visual navigation systems are based on a combination of computer vision, deep learning, and multi-sensor integration. The core of such systems consists of visual odometry (VO) and visual simultaneous localization and mapping (Visual SLAM) methods.

Comparative experiments of VO methods on the standard KITTI dataset confirmed the effectiveness of modern optimization and dynamic models. Research into Visual SLAM algorithms in dynamic scenes showed a significant advantage for methods implementing loop closure mechanisms and filtering of moving objects. At the same time, it should be emphasized that the provided numerical evaluations were mostly obtained in local test ranges and small environments with good visual landmarks (urban scenes, laboratory, or corridor routes). The conducted review demonstrates existing and promising solutions and outlines directions for the further development of this technology.

Approach to implementation of the accelerated zero knowledge identification of remote abonents

2026-04-27T03:30:20+03:00

The article proposes a method of accelerated zero knowledge identification, aimed at use in systems that interact with a large number of remote abonents. Reduction of the system identification time of a remote abonent is achieved due to the organization of an asymmetric computing load on the system and the abonent in the process of implementing the proposed procedure of cryptographically strong identification.

Formalized procedures for abonent registration and identification before the start of a remote interaction session with the system have been developed. The work of the procedures is illustrated by examples.

It is theoretically proven and experimentally confirmed that the proposed approach allows to reduce the time of abonent identification by the system by 2-3 orders of magnitude compared to known methods of cryptographically strong identification.

Infrastructure for integration of software and hardware complexes of special-purpose information systems

2026-04-27T03:21:40+03:00

This article examines modern approaches to building an integration infrastructure for hardware-software complexes of special-purpose information systems. Architectural models of integration are analyzed. A reference architectural model of integration is proposed, tailored to the conditions of modern combat operations and the degradation of communication lines. An integration infrastructure model is proposed that includes a set of support services for information and communication systems: situational awareness, operations planning, tasks and orders, operational information, and spatiotemporal information. The support services for information and communication systems define the principles for building IT services and their interrelationships. Additionally, requirements for the creation of IT services are derived from each support service. The following model of integration infrastructure for hardware-software complexes of special-purpose information systems is proposed: information systems and communications support services define a set of IT services; IT services are provided to three groups of specialized clients who use specific software-hardware complexes; IT services and software-hardware complexes are linked by 5 implementation scenarios for heterogeneous information transmission networks; the integration of IT services is determined by 5 information and communication systems support services. A method has been developed to optimize the process of providing integration infrastructure IT services to specific categories of clients, which are offered to three groups of specialized clients using specific hardware and software systems. The method differs from others in its clearly defined algorithm for using five implementation scenarios for heterogeneous data transmission networks, which ensures data transmission under degraded communication line conditions and minimizes the processing time of client requests to information and optimize the performance of the integration infrastructure under a specified load for the selected system configuration.

DDoS attack detection methods in software-defined networks (SDN): a comparative analysis of classical and modern machine learning and deep learning approaches

2026-04-27T03:16:03+03:00

This paper examines DDoS attack detection methods in software-defined networks (SDN). The aim of the study is a comparative analysis and systematization of classical and modern approaches to attack detection in SDN environments.

Traditional protection methods are analyzed, including entropy-based statistical analysis, threshold and rule-based systems, and conventional machine learning algorithms, with their limitations under modern adaptive attacks identified. Modern machine learning and deep learning approaches are investigated, including ensemble methods, hybrid architectures, adversarially robust models, and federated learning, along with their respective advantages and drawbacks. The accuracy and efficiency of various methods are compared on the basis of current published research results.

The paper proposes further investigation of two-tier machine learning and deep learning architectures and federated learning in software-defined networks, with the aim of improving attack detection accuracy, adversarial robustness, and adaptability to changes in the network environment. Results of the method comparison are presented with emphasis on their effectiveness in real SDN deployments.

Methods for evaluating the energy efficiency of large language models

2026-04-27T03:08:31+03:00

This paper examines methods for evaluating the energy efficiency of autoregressive large language models based on the transformer architecture, focusing on representatives of the Cogito, Phi-4, Mistral, and RNJ-1 families. Given the rapidly increasing computational complexity of attention mechanisms and the associated power demands during inference, the study emphasizes experimental measurement of model power consumption on a consumer-grade NVIDIA RTX 3070 Ti GPU using CUDA acceleration. The proposed approach enables quantitative assessment of average, minimum, and maximum power draw, as well as comparative analysis of relative energy efficiency across models in typical text-generation scenarios. The obtained results provide a baseline for further research on energy-efficient deployment of artificial intelligence systems and highlight the industrial and societal importance of reducing the energy footprint of modern LLMs. In addition, the article presented a number of other approaches to improving the energy efficiency of LLM, such as query routing and dynamic power adjustment when working with query decryption. The comprehensive use of various optimization methodologies is an important factor in the development and implementation of neural networks LLM.

Adaptive leader election method in mobile distributed systems based on dynamic trust integration

2026-04-27T02:32:52+03:00

The development of mobile distributed systems, particularly Flying Ad-hoc Networks (FANETs), requires a transition to decentralized control architectures involving local leader election. Under hostile conditions, classical additive clustering methods (e.g., Trust-WCA) reveal vulnerabilities to compromised Byzantine nodes (participants that intentionally act destructively while simulating legitimate behavior) exhibiting high radio channel performance. This paper proposes an adaptive leader election method based on a non-linear multiplicative function. The application of a dynamic trust exponent, regulated by an LSTM neural network module depending on the threat level, acts as a filter ("soft gate"), preemptively isolating suspicious nodes. Based on Trace-Driven Simulation results, a relative decrease of 11.6% in the frequency of electing compromised leaders was proven, alongside a simultaneous 0.8% improvement in the network's topological efficiency.

Improving remote control efficiency for dynamic systems through criterial evaluation of visual human-machine interfaces

2026-04-27T02:24:29+03:00

The scientific article addresses the critical socio-technological challenge of enhancing the efficiency of remote control for dynamic objects and ensuring the ergonomic quality of visual human-machine interfaces. Within the scope of the study, an information technology for evaluation has been developed, based on a hierarchically structured system of criteria. This technology provides a formalized analysis of human-machine interaction efficiency by determining the integral quality index of the interface. It is demonstrated that the efficiency of remote control systems is significantly determined by the balanced distribution of interface characteristics across the criteria of cognitive workload and functional completeness. This enables a justified selection and optimization of interface solutions tailored to specific tasks and mission types, thereby increasing the performance of operator activities within complex information environments.

Principles of using AI agents in the construction of software code templates

2026-04-27T01:56:48+03:00

The paper is devoted to the study of principles for using AI agents in the construction of software code templates under conditions of increasing complexity of modern software systems. The relevance of the topic is determined by the limitations of traditional code generation approaches, such as the lack of contextual consistency, instability of results, and insufficient quality control. The study analyzes modern approaches to programming automation, including template-based methods, neural network models, and large language models (LLMs), and also examines the concepts of multi-agent systems in software engineering.

A formalization of the code template generation problem is proposed as a mapping of a set of requirements, context, and technology stack into a structured software template. A concept of a multi-level agent-based system is developed, which includes analysis, planning, generation, validation, and optimization agents interacting within an iterative process.

A system of principles for constructing such a system is formulated, including the principles of contextual awareness, multi-level generation, iterative validation, modularity, and integration with DevOps processes. The architecture of the system and the algorithm of its operation are proposed, taking into account direct, feedback, and contextual data flows.

To evaluate the effectiveness of the proposed approach, a formalized multi-criteria model is developed, enabling theoretical comparison with alternative solutions. The obtained results can be used in the development of intelligent programming support systems and the automation of software development processes.

An agent-based model for the early detection of trust anomalies in industrial IoT networks

2026-04-26T20:48:23+03:00

The paper proposes an agent-based model for the early detection of trust anomalies in Industrial Internet of Things networks operating under dynamic conditions. The relevance of the study is determined by the need for timely identification of abnormal node behavior that may indicate communication instability, compromised operation, false data injection, or data inconsistency before explicit failures occur. The proposed model is based on the interaction of local monitoring agents, trust evaluation agents, and a coordination agent, which together provide distributed observation of node behavior and cooperative interpretation of suspicious deviations. A compact set of trust-related indicators is used, including data consistency, communication stability, packet delivery regularity, response delay, and interaction reliability. To formalize anomaly detection, an integrated trust score and a threshold-based decision rule are introduced. An illustrative synthetic scenario is presented to demonstrate the operation of the proposed model under normal, degraded, and anomalous node states. The results show that gradual degradation of trust-related indicators leads to a decrease in the trust score and enables early anomaly signaling before explicit system-level failure becomes evident. The proposed approach can be used as a conceptual basis for distributed trust monitoring in dynamic Industrial IoT environments.

Capacity planning method for resource consolidation in mass cloud web hosting

2026-03-06T12:52:32+02:00

In this scientific article, a capacity planning method is proposed for resource consolidation in the mass cloud web hosting segment, based on mathematical modeling of aggregated stochastic loads and queueing theory. The method optimizes the allocation of RAM for low-traffic resources by consolidating traffic into a single computing pool while strictly maintaining the target reliability level (SLA 99.9%). Through simulation of the system’s weekly operating cycle, it is shown that the proposed method provides an integrated resource saving of 16%, mitigates the impact of stochastic load peaks, and reduces the provider’s operational costs without degrading service quality.

Comparative analysis of swarm intelligence metaheuristic algorithms for group UAV control

2026-03-06T12:49:19+02:00

This article is devoted to conducting a comparative analysis and evaluating the efficiency of modern metaheuristic algorithms introduced no later than 2021-namely, African Vultures Optimization Algorithm (AVOA), Harris Hawks Optimization (HHO), Dung Beetle Optimizer (DBO), and Snake Optimizer (SO) which can be applied to decentralized unmanned aerial vehicle (UAV) swarm control tasks. The research is timely and driven by the shift in UAV management from single-unit control to the coordination of autonomous swarms capable of executing labor-intensive tasks while minimizing risks inherent to centralized systems, specifically by reducing critical communication latency and enhancing resilience against control node failures.

The research methodology was implemented using simulation modeling in a Python version 3.12.9 environment. The testing suite consisted of 23 classic benchmark functions, on which the four aforementioned algorithms were evaluated. Accuracy (Mean), iterative stability (Std), and convergence speed at a fixed dimensionality of D = 30 for each algorithm were selected as the primary comparison metrics.

The results of the numerical experiments demonstrated a clear specialization for each algorithm: it was established that AVOA possesses the highest, near-benchmark localization accuracy and exceptional stability across most benchmarks compared to its competitors. This makes the algorithm the optimal choice for swarm targeting or spatial station-keeping tasks. DBO proved to be the superior algorithm for global exploration in complex terrains, demonstrating the highest efficiency in path planning through unknown areas with a high density of obstacles. The HHO algorithm showed sufficiently balanced indicators of accuracy and stability, particularly on curved landscapes. Generally competitive with AVOA, HHO is also excellently suited for target tracking tasks. The SO algorithm proved to be the least effective on average due to slow convergence, yet it maintains sufficient effectiveness for use in low-dimensional tasks.

Assessment of service traffic dependences in Ad-hoc networks

2026-03-06T12:45:48+02:00

The article devoted to the problem of excessive service traffic in wireless ad-hoc networks, which arises due to the lack of a fixed infrastructure, dynamic topology, and limited node resources. Service traffic is interpreted from an information-theoretic point of view as a mechanism for reducing uncertainty about the network state. A mathematical model is proposed that relates the intensity of service traffic to the number and density of nodes, the level of interference, and the signal-to-noise ratio in communication channels. It is shown that for a fixed deployment area, an increase in node density leads to a superlinear, close to quadratic, increase in network management overhead. Special attention is paid to the possibilities of reducing the effective node density due to algorithmic routing mechanisms, in particular, approaches such as ND-AODV, which allow significantly reducing the volume of service traffic in dense networks. The results obtained can be used to assess the scalability of ad-hoc networks and develop more effective management protocols.

Method of multi-modal parametrical robust optimization of digital flight control systems

2026-03-06T12:41:26+02:00

In this paper it is proposed a method for achieving a trade-off between performance and robustness in the design of digital control systems for dynamic plants described by nominal and parametrically perturbed models. The design process accounts for both deterministic and stochastic cases. The developed method is based on the application of a multi-model Н₂/Н_ robust optimization approach specifically for digital systems. The robust parametric optimization procedure is conducted in an interactive mode until the system's performance and robustness meet the specified requirements. The effectiveness of the developed multi-model parametric robust optimization method for digital flight control systems is validated through the synthesis of a digital heading stabilization system for a small UAV.

Overview of methods for determining the location of sound sources

2026-03-06T12:37:17+02:00

The paper provides an overview and systematic comparison of modern methods for sound source localization, including time-based (TDoA, GCC), subspace-based (MUSIC, ESPRIT), beamforming methods, statistical tracking approaches, and neural network and hybrid solutions. For each class of methods, the key advantages and limitations are analyzed, as well as their suitability for use in real acoustic conditions, in particular in the presence of noise, reverberation, and multiple simultaneous sources. Based on a comparative analysis, it is concluded that hybrid architectures that combine physically interpreted features with adaptive machine learning models are the most promising for practical systems. A hybrid concept of multi-source localization is proposed, which uses GCC cross-spectral features and a multi-head neural network model to estimate the number of active sources, their directions, and the confidence of the prediction.

Improving the reliability of drone navigation solutions based on the interacting multiple model Kalman filter

2026-03-06T12:30:25+02:00

The article considers the problem of decision-making in drone navigation in noisy conditions and with sensor degradation, when the classic Kalman filter with fixed noise statistics is unable to adapt quickly enough to changes in measurement quality. An approach based on the Interacting Multiple Model Kalman Filter is proposed and implemented, in which several estimation modes with different assumptions about the degradation of LiDAR, optical flow, IMU, and compass work in parallel, and the resulting state estimate is formed by a weighted combination of the posterior probabilities of the modes.

The evaluation was performed in a drone flight simulation in two series of experiments. In series A, a comparison of KF and IMM-KF was performed at four noise levels, where the success rate, mission duration, and RMSE of the state estimation on successful runs were measured, as well as the causes of failures. In series B, the correctness of the IMM mode switching mechanism was verified in controlled scenarios of spatial and temporal degradations. The probabilities of modes, the determination of the noise mode by the drone, the contrast between active and inactive intervals, and the switching delay were analyzed. The results show that IMM-KF increases stability in strong degradations without changing the autopilot, and the mode probability metric is an interpretable indicator of changes in the quality of sensor measurements.

Mapping cyclo-dynamic dataflow into pipelined datapath

2026-03-06T12:24:44+02:00

The paper discusses the relevance of high-level synthesis (HLS) systems in designing pipelined datapaths. The goal of the research is to explore methods for mapping algorithms into a pipelined datapath that implements cyclic data flow graphs with dynamic schedules. The proposed method involves creating and optimizing cyclo-dynamic data flow graphs (CDDFG) and describing them in VHDL. A set of rules for arranging the correct CDDFG and respective finite state machine is proposed. The derived CDDFG is mapped into the pipelined datapath, as well as the synchronous dataflow. The method demonstrates its effectiveness through examples of run-length encoding decompression and LZW file decompression devices, which are implemented in field programmable gate arrays. The proposed method can be used manually or be implemented in HLS tools.

Methodology of multi-agent system for misinformation and disinformation detection

2026-03-06T12:19:43+02:00

The rapid growth of generative artificial intelligence has led to an unprecedented surge in disinformation, significantly complicating traditional media monitoring methods and necessitating the implementation of new automated verification tools. However, the effectiveness of existing manipulation and disinformation detection systems often relies on powerful commercial Large Language Models (LLMs), which entail high operational costs, limited accessibility, and risks associated with transferring confidential data to external servers. This paper assesses the potential of utilizing Small Language Models (SLMs) as a foundation for building autonomous systems to counter information manipulation. A proposed approach to constructing a verification system based on multi-agent interaction is presented, in which a complex task is decomposed into subtasks for specialized agents: a manipulation detector, a narrative analyst, and a fact-checker. The results of an experimental comparison of local and cloud-based models indicate that the proposed approach is more secure and cost-effective, achieving high analysis accuracy while maintaining full control over user data.

Intelligent traffic engineering methods in SDN: Performance analysis and prospects for application

2026-03-06T12:14:45+02:00

Software-defined networks (SDN) create favorable conditions for deploying intelligent traffic management methods thanks to centralized control and global visibility of the network state. At the same time, traditional load balancing algorithms cannot effectively adapt to dynamic traffic changes and complex topologies. This paper analyzes the effectiveness of the intelligent load-balancing method LBBNN, based on an artificial neural network that considers the global network state when selecting a route. The evaluation was carried out in a Mininet environment on sparsely connected, fully connected, and de Bruijn topologies. The results demonstrate reduced delays and packet loss and a more uniform load distribution compared to Round Robin, Dynamic Load Balancing (DLB), and modified ECMP. The neural-network approach improves quality of service (QoS) and enables more efficient use of network resources in dynamic SDN environments.

Image compression method based on multi-scale decomposition with increased weights in spline layer nodes

2026-03-06T12:10:24+02:00

The article is devoted to the development and improvement of image compression methods. Information compression is critically important in the modern digital world, as the volume of data is constantly growing, and effective management of this data is becoming more and more relevant. To solve this problem, algorithms have been created that allow performing a compression operation with the loss of a certain amount of information (Lossy) or without loss (Lossless). This article focuses on the development of a lossy compression method, where less noticeable data is removed, which gives a high compression ratio, but leads to a decrease in quality (for example, JPEG, fractal, some MPEG methods). The main task of scientists is to find such an algorithmic solution that will allow, at a higher compression ratio, to leave the highest possible image quality. The main mathematical apparatus that provides lossy compression is the discrete cosine transform (DCT) or wavelet methods for converting pixels into the time domain. The authors proposed to use a different approach, namely, in the process of multiscale decomposition, to use spline functions as the basis function instead of wavelets. Accordingly, a spline multiscale decomposition with increased weights in the spline gluing nodes was developed. The advantages of such an algorithm in comparison with the direct procedure of spline multiscale decomposition and with known image compression standards were also evaluated.

Method for accelerating the calculations of exponent on Galois fields by combining parallel processing and pre-calculations

2026-03-06T12:04:00+02:00

Method for exponent calculation on a Galois Field has been developed and researched, with its parallel implementation on an h-core terminal microcontroller for cryptographic applications. The acceleration was achieved due to the following factors: distribution of calculations between h-cores, fast exponentiation to the power of 2^k, and simultaneous processing of k digits of the exponent code through the complex use of precomputations. A mathematical proof of the proposed method is given, as well as procedure its operation whose are illustrated be numerical examples.

It has been theoretically proven and experimentally confirmed that on a four-core processor and a fragment size of twenty bits, a 20.34-fold acceleration is achieved compared to the classical method of exponentiation on finite Galois fields.

Experimental study of the efficiency of scalable YOLOv8(n-x) architectures in multi-object tracking tasks based on the ByteTrack algorithm

2026-03-06T11:46:53+02:00

The article is devoted to the results of research on the performance of a computer vision system for intelligent road traffic monitoring tasks. The problem of finding a balance between Inference Speed and Tracking Stability in dense urban traffic conditions is investigated.

Classic approaches to multi-object tracking (MOT) often face the problem of object loss during partial overlap (occlusion). The modern ByteTrack algorithm offers a solution to this problem by using two-level association of detections with existing tracks, which makes it particularly resistant to overlaps. This algorithm is the optimal solution for real-time systems, as it minimises computational costs while maintaining the highest trajectory retention quality among the considered analogues. The basic principle of the IoU tracker, which underlies the geometric component of ByteTrack, was considered. The paper experimentally proves that for real-time urban monitoring tasks on mid-range hardware (NVIDIA RTX 2060 level), the YOLOv8m model is the best choice. The paper provides a detailed analysis of the five tested YOLOv8m architectures: Nano, Small, Medium, Large, and X-Large. Through empirical comparison of the performance and quality of models of different scales based on real video data from a traffic intensity monitoring system, the optimal neural network architecture of the YOLOv8 family for integration with ByteTrack was determined.

As a result of the analysis of modifications of the YOLOv8 neural network in combination with the ByteTrack tracker, a conclusion is drawn regarding a significant non-linear dependence between the computational complexity of the neural network detector and the final quality of multi-object tracking.

A method to determine the criticality of gateways in a LoRaWAN network

2026-03-06T11:33:27+02:00

LoRaWAN networks rely on gateways to relay packets from resource-constrained end devices to the network server, creating a potential single point of failure in network infrastructure. While traditional graph centrality metrics exist for general networks, they fail to capture LoRaWAN-specific characteristics such as adaptive data rate (ADR) mechanisms, spreading factor orthogonality, and asymmetric node-gateway connectivity patterns. This paper introduces a gateway criticality metric specifically designed for LoRaWAN networks, as well as an algorithm that employs this metric in an analysis task in order to identify and prioritize critical infrastructure components before failures occur.

The metric for gateways presented in this study combines three LoRaWAN-specific factors: connected node count (coverage contribution), exclusive node count (single points of failure), and served traffic volume (application-level importance). Unlike traditional centrality measures, the algorithm presented accounts for ADR's ability to adaptively increase spreading factors when gateways fail, recognizing latent redundancy that becomes accessible during outages. The exclusive nodes are weighted most heavily (50% contribution), as their isolation has immediate, unavoidable impact on network availability.

The metric is validated through 300 simulations across diverse network topologies spanning 7 to 50 nodes and 1 to 5 gateways. Results demonstrate strong correlation between criticality scores and measured failure impact. High-criticality gateway failures (C ≥ 0.7) caused packet delivery ratio (PDR) drops of around 60%, while low-criticality failures (C < 0.3) produced up to 25% drops. The metric generalizes across reviewed network scales and topologies, from simple dual-gateway deployments to complex scenarios with 50 nodes and 5 gateways.

Risk assessment and information security ensuring of Starlink satellite communication channels at aviation enterprises

2026-03-06T11:18:16+02:00

The article investigates the specific features of risk assessment and information security ensuring of Starlink satellite communication channels when integrated into aviation enterprises as critical infrastructure objects. A structural analysis of the Starlink architecture was conducted, and key threat vectors in the space, ground, and network segments were identified. Based on STRIDE and ISO/IEC 27005 methodologies, an adapted threat model for aviation enterprises was developed. A formalized approach to quantitative risk assessment considering segment architecture, asset criticality and security control effectiveness is proposed. Conceptual principles of a multi-layered protection system are substantiated, including network segmentation, additional cryptographic protection, multi-factor authentication and integration with security monitoring systems. The results can be applied during the integration of satellite communication channels into aviation enterprise information infrastructures.

Container model of a multi-component system in the digital twin architecture

2026-03-06T10:37:20+02:00

The paper addresses the problem of formal modeling of multi-component computer systems within the architecture of digital twins. The relevance of the study is обусловлена the rapid development of container technologies, microservice architectures, and distributed computing environments where systems consist of numerous interacting components. A container-based model of a multi-component system is proposed in which a container is interpreted as a structural and functional subsystem with a defined state space, dynamic operator, and resource characteristics. The system is represented as a set of interacting containers, for which the interaction mechanisms, local component dynamics, and resource constraints are formally defined. An aggregated model of the global system dynamics is constructed, enabling system-level analysis of behavior. Within the digital twin architecture, a mapping between the physical system and its digital representation is introduced, and conditions of structural, dynamic, and resource consistency are formulated. The proposed approach integrates container-based decomposition with a mathematical representation of system behavior and can serve as a formal basis for digital twin architectures in distributed, cloud, and edge computing environments.

Software architecture of the system for expert assessment of the efficiency of unmanned aviation systems

2025-12-05T12:26:15+02:00

This paper proposes a software architecture of an automated system for expert assessment of the quality of unmanned aviation systems, designed to collect individual expert assessments, their statistical processing and generation of interactive reports and recommendations for improving the ergonomic characteristics of UAS. The program allows you to systematize expert assessment of the quality of unmanned aviation systems when direct measurements are impossible or ineffective. The program is implemented on the basis of a cross-platform Web solution with a client part in HTML5/JavaScript and server processing in PHP, the modular structure of the system is given (user account, project creation, expert evaluation module, data analysis, reports and recommendations), approaches to data storage and protection, authentication mechanisms and an algorithm for aggregating results based on a weighted average are described. Functional capabilities for dynamically adjusting weights and thresholds are proposed, as well as tools for ensuring the integrity and accessibility of examination results. The final part of the work provides an example of software implementation and its interface.

Comparative analysis of convolutional and multilayer perceptron neural networks for resource allocation in distributed computing systems

2025-12-05T11:58:23+02:00

Efficient resource distribution in heterogeneous distributed computing systems requires intelligent node selection mechanisms capable of adapting to dynamic system conditions. This research presents a comparative evaluation of Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) architectures for binary node suitability classification in distributed task scheduling environments. The study employs five synthetic datasets ranging from 100 to 2000 instances, with each node characterized by four critical attributes: performance, security level, baud rate, and reliability. Experimental results demonstrate that the MLP architecture achieves validation accuracy between 91% and 98.8% with high consistency across dataset sizes, while the CNN architecture shows fluctuating performance ranging from 85% to 94%. The key finding reveals that MLP architectures outperform CNNs for tabular node data due to better alignment with unstructured attribute relationships, as fully connected layers naturally handle unordered features without imposing spatial locality assumptions. This empirical analysis provides practical guidance for architecture selection in AI-based resource schedulers, demonstrating that simpler architectures can outperform more complex ones when appropriately matched to problem structure. The findings contribute evidence-based recommendations for distributed systems practitioners implementing neural network-based scheduling solutions.