AUTOMATIC CONTROL AND COMPUTER SCIENCES最新文献

The mining algorithm in smart city blockchain systems using the Proof-of-Work consensus mechanism is studied. Well-known studies in the field of selfish mining detection are analyzed. A method for protecting a blockchain from selfish mining attacks is presented, and a selfish mining detection plugin is developed based on this method. This plugin is designed for miner software and enables the analysis of data patterns received from the mining pool. The proposed solution outperforms existing selfish mining detectors by identifying the attacking mining pool and has lower error rates.

This paper presents a study of the application of large language models (LLMs) to predict events based on LLM agents—autonomous systems that use LLMs for reasoning, decision making, and interaction with the environment. Various architectures of LLM agents are analyzed: cooperative systems (ChatDev, MetaGPT), multiagent debates (MAD, ChatEval), agents for web tasks (WebAgent, WebVoyager), and simulation agents (Generative Agents, EconAgent). Particular attention is paid to the features of predictive modeling based on LLMs, where classical approaches (regression, time series) are replaced by agent-based modeling and predictive engineering. This article presents the results of an experiment on predicting the outcome of a selected conflict using an LLM agent (Mistral, DeepSeek) and the retrieval-augmented generation (RAG) approach based on data from analytical agencies, opinion leaders, and news sources. The convergence of forecast estimates of polarized sources is revealed and requirements for forecasting systems are formulated: weighting sources according to expert significance, filtering neutral data, and sample balancing. Requirements are put forward for the selection of data assessed by simulation LLM agents.

The problem of optimizing neural networks for large language models (LLMs) such as ChatGPT is discussed. One of the directions being developed for optimizing LLMs is knowledge distillation—the transfer of knowledge from a large teacher model to a smaller student model without significant loss of accuracy of the result. The existing methods of knowledge distillation have certain disadvantages: inaccurate knowledge transfer, long learning process, and error accumulation in long sequences. A combination of methods that contribute to improving the quality of knowledge distillation is considered: selective teacher intervention in the student’s learning process and low-rank adaptation. The proposed combination of knowledge distillation methods can be applied to problems with limited computational resources.

A study of adversarial attacks on classical machine learning (ML) algorithms in the context of network threat detection is presented. An overview of ML models that are used to perform various tasks in computer network security systems is presented. A formal description of the threat model is provided, as well as a classification of adversarial attacks. The network traffic of the WEB-IDS23 dataset was classified using classical machine learning models: k-nearest neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM). Adversarial attacks such as the fast gradient sign method (FGSM), projected gradient descent, Carlini & Wagner (C & W), and DeepFool are implemented on these ML algorithms. The impact of the adversarial attacks implemented on listed classical machine learning algorithms is analyzed.

The architecture of large language models (LLMs) and the possibility of their application to automate fuzz testing software are studied. As a result of the study, a method is developed that enables fuzz testing applications with a graphical interface in the Windows operating system. A software prototype implementing the proposed method is developed and tested.

The existing approaches to building active auditing systems for Internet-of-Things (IoT) devices are studied. A robust approach to auditing IoT devices using fault-tolerant distribution is proposed. A comparative analysis of consensus-achieving algorithms in distributed systems and means of implementing active auditing is conducted.

A systematic analysis of methods for evaluating the effectiveness of fuzzers is conducted. A minimal but complete set of metrics is identified: branch coverage, number of unique crashes, and time to first crash. A normalized integral indicator is proposed that allows post hoc comparison of the results of different tools without reruns.

This paper studies the problem of reducing the attack surface from an internal attacker in heterogeneous systems for processing and storing big data by selecting the optimal data obfuscation method based on anonymization technologies. The study analyzes the terminology and systematizes data-hiding methods to reduce the attack surface in big data processing and storage systems. A formal formulation of the problem of finding the optimal data obfuscation method and an algorithm for solving it across various types of datasets are proposed, taking into account evaluation criteria specific to each class of methods. The implementation of a software prototype for supporting decision-making and selecting the optimal method for solving practical problems is described. Experimental testing and analysis of its results are carried out.

The main biometric characteristics reflecting changes in the psycho-emotional state of a user of an information system are examined. They are ranked using the pairwise comparison method, as a result of which voice and keystroke dynamics are identified as most suitable for further research. Criteria for the preliminary identification of potential internal information security violators based on changes in the considered biometric characteristics are defined. A convolutional neural network model is developed and tested to solve the stated problem.

The principles of construction and operation of honeypot systems are studied. Existing detection methods are analyzed, and their advantages and disadvantages are highlighted. A detection method based on the analysis of command execution delays is proposed. A universal detection method based on combining the results of the methods is proposed. A software prototype of the detection system is developed and tested.