Information and Control Systems

Traffic sign recognition through the use of an Internet of Things system and deep learning

Thu, 01 May 2025 12:35:43 +0000

Introduction: Automatic traffic sign recognition is an important component in modernizing traffic safety systems and minimizing related incidents. Purpose: To develop a Vietnamese traffic sign detection system that integrates IoT technology, cameras, sensors, and deep learning techniques. Results: In this study, we introduce an advanced computer vision solution that uses the neural architecture of YOLOv8 to recognize and classify traffic signs in Vietnam in real time. In addition, the system integrates the detected sign locations into Google Maps and provides a comprehensive database at all locations in urban areas. The data collection process is implemented by first capturing real-life images on the road and then combining them with the existing Vietnam traffic sign dataset from Kaggle data. To improve the reliability of the dataset, Mosaic data augmentation technique is applied. For real-time traffic sign recognition, a Raspberry Pi 4 board displays detected road signs on an HMI screen, while a custom-developed mobile application detects, recognizes, and notifies users of their locations. The trained system achieves a mAP50 value of 89% and an accuracy of 90%, which is pretty good for traffic signs detection. Moreover, the YOLOv8 model gives a precision and recall value of 89.7% and 86.8%, respectively, which is considerably higher as compared to the SVM and MSER/HOG methods. Practical relevance: We have succeeded in creating a real-time road sign recognition system integrated into Google Maps to meet the needs of traffic control, safety and accident reduction in the streets of Vietnam.

Application of spatial approach to the problem of archiving measurement data

Thu, 01 May 2025 12:35:43 +0000

Introduction: In the case of telemetry data, which has received little attention in recent years, there are a number of both specialized and general-purpose algorithms available for processing it, yet none of these have demonstrated even acceptable performance and have a number of fundamental limitations. Purpose: To develop an efficient algorithm for compressing/restoring measurement data frame sets based on a spatial approach to their representation, and a format for storing their compressed representation for the archiving task. Results: We propose an algorithm for compressing measurement data frame sets which are three-dimensional difference-bit matrices, and consider various methods for displaying them in the body of a fractal cube. To store a compressed data representation, we develop a specialized description format that can be implemented effectively in the task of archiving them. We give the estimates of the speed and quality of the developed algorithm, the efficiency of which is compared with both the algorithm based on the construction of Huffman codes and the DEFLATE algorithm, which is part of the WinRAR utility, the latter, in its turn, being the de facto industry standard for solving compression and archiving problems. We prove that the developed algorithm is significantly superior to the Huffman and DEFLATE algorithms in both speed and compression ratio. Practical relevance: The use of the proposed algorithm can lead to significant technical and economic benefits for industrial enterprises and improve the efficiency of telemetry data transmission systems. Discussion: The main disadvantage of the developed algorithm is the nonlinear dependence of the compression time on the dimension of the compressed data, which is a consequence of both the impossibility of effectively segmenting the data for their separate processing and the need for multi-pass processing. Solving these problems opens up a number of promising directions for further development of the algorithm.

Classification of group control precedents

Vyacheslav Konstantinovich Abrosimov, Ekaterina Sergeevna Mikhailova — Thu, 01 May 2025 12:35:43 +0000

Introduction: In the context of the development of autonomy for unmanned vehicles, it is important to improve the methods for increasing the level of autonomy. One of such methods is the use of precedents that allow the system to make decisions in the current situation, based on past experience. Purpose: To develop a classification of group management precedents using the hierarchical and faceted method for the effective use of precedents, to estimate the capacity and fill factor of the classification of such precedents. Results: We introduce new definitions of a precedent state, precedent collision, precedent situation and precedent facts. We confirm the applicability of common classification methods – hierarchical and faceted ones – to group management precedents. We present the depth of the hierarchical classification for various classes that is sufficient for solving practical problems. We introduce two new definitions: “core” and “fringe” of a precedent, which are required to build a faceted classification. They make it possible to select several variants of typical precedents, in which the states of the group can described only by the “cores” of features, without paying special attention to the “fringe”. Using the logical-linguistic model as a base we create a facet formula for precedents, taking into account the “core” and “fringe” of the precedent. Finally, we carry out an assessment of the capacity and fill factor of the precedent classification. As a result, the capacity for each classification is determined. The analysis of the features of precedents allows us to conclude that the occurrence of a precedent in group management is a fairly complex phenomenon where various aspects of the precedent are taken into account. On the basis of the classification of precedents, a definition of a group precedent is formulated. Practical relevance: The classification of precedents makes it possible to implement more efficient control of unmanned vehicles, as it facilitates the search for suitable precedents when it is necessary to make a decision during the mission.

Active protection method against cyberattacks for the objects of critical information infrastructure based on the interruption of the process of an intruder’s impact

Thu, 01 May 2025 12:35:43 +0000

Introduction: The development of IT-technologies and the specifics of dynamic interaction of the parties within the conflict between critical information infrastructure objects and intruders lead to the emergence of new cyberattacks. Purpose: To develop a new approach to monitoring, analyzing and interrupting the attack chain even before it achieves the goal of invasion at early stages of attacks, with the results of confrontation modeling taken into account. Results: We structure the process of implementing a cyberattack, including the main phases: analysis and implementation, active impact and completion with data output. The features of modern methods of multi-stage attacks and the programs used by an intruder are taken into account. We develop a time model of a multi-stage attack and a state graph of an intruder's actions, which makes it possible to calculate the probabilistic and time characteristics of a successful intruder's impact on the network. We propose an algorithm for interrupting a cyberattack at the network scanning stage, which involves identifying attempts to scan the network, collecting and processing information about the intruder, as well as forming countermeasures to interrupt the attack and implementing an automated control of the system's effectiveness with the ability to adjust protection measures. As a result, we model the process of implementing active protection against cyberattacks for the information and computing network, which demonstrates the effectiveness of the proposed protection method. Practical relevance: The use of the method of interrupting the cyberattack cycle at critical information infrastructure objects will increase the efficiency of suppressing the impact of a cyberattack at early stages of penetration.

Automated system for analyzing weakly structured threat intelligence data using large language models

Thu, 01 May 2025 12:35:43 +0000

Introduction: The lack of efficiency and the complexity of the analysis and management of semi-structured and structured Threat Intelligence data collected from various sources is a pressing issue. Purpose: To increase the efficiency (which implies reducing the labor costs of experts and ensuring the completeness and promptness of the knowledge base updates) of Threat Intelligence data analysis using artificial intelligence methods. Results: The conducted analysis of approaches to the construction of Threat Intelligence data management systems has shown as a promising direction the use of large language models as part of a “question-answer” decision support system with an extended augmented generation mechanism. We develop a structural diagram and functional model of a data mining system for the Threat Intelligence data. It is based on large language models using a retrieval augmented generation pipeline. In addition, we develop an algorithm for semantic tagging and extracting information from weakly structured data. We present a research prototype (component model, microservice architecture) of the software. A distinctive feature of the system is the integration of information from multiple sources using a retrieval augmented generation pipeline. A computational experiment on the prepared data set has shown the consistency of expert responses for the set of collected documents with the responses proposed by the system at the level of 3.98 points out of 5. The BERTScore F1 metric value is 0.89, and the RAGAS Answer Correctness metric value is 0.822. Practical relevance: The use of large language models provides the accumulation of knowledge about current attack scenarios within a single knowledge base, which makes it possible to enhance the efficiency of data processing to enable a significant support to information security specialists during the analysis of current threats, vulnerabilities and cyber attack scenarios by reducing labor costs (time for collecting and processing), and thereby to increase the security of corporate information systems. Discussion: A further improvement of the efficiency of the Threat Intelligence data analysis is possible on the basis of constructing heterogeneous “expert committees” out of large language models and multi-agent systems.

Decoding of single error bursts using minimal burst length criteria and information sets

Maria Nikolaevna Isaeva — Thu, 01 May 2025 12:35:43 +0000

Introduction: To improve the characteristics of data transmission and storage systems the specificity of noise should be taken into account, which tends to form the error bursts in many practical situations. Information set decoding for independent errors is the best decoding approach for random linear codes, but has exponential complexity. Purpose: To analyze the error-correcting capability and computational complexity of algorithms based on information sets in order to correct single error bursts. Results: We suggest the decoding criteria of minimal burst length for single bursts correction using information sets. We prove the correction of all error bursts within burst-correction capability of the code using this criterion. The task of decreasing the number of information sets required for correct decoding is considered, for this task the methodology of constructing the information sets with limited diameter, which are called dense sets, is investigated. We perform the analysis showing the possibility of decreasing the number of information sets from O(n) to O(1). We estimate the probability of information sets construction and formulate the optimization problem for maximizing this probability. Practical relevance: The developed decoder of single error bursts basing on minimal burst length criteria, as well as methodology of construction the collection of dense information sets with cardinality O(1) guarantee the realization of burst-correction capability of random linear codes. Discussion: The results obtained show that the problem of single error bursts correction with random linear codes is polynomial. At the same time, the proposed methods make it possible to get a high probability construction of information sets for codes with a relatively low rate, and these probabilities can significantly change due to the use of more practical classes of codes, which may be considered as future research directions.

Information about the authors

Unknown — Thu, 01 May 2025 12:35:43 +0000