9004 2071-8217 3 2025 1-213

RAR RUS 9-19 0000-0001-8206-2915 Peter the Great St. Petersburg Polytechnic University Ivanov Denis ivanov@ibks.spbstu.ru 0009-0002-2463-3876 Gavrichkov Daniil gavrichkov.da@edu.spbstu.ru

Peter the Great St. Petersburg Polytechnic University

0000-0002-2009-5460 Peter the Great St. Petersburg Polytechnic University Ovasapyan Tigran otd@ibks.spbstu.ru Fuzzing of closed-source software using large language models The architecture of large language models and their potential application for automating fuzzing testing of software are investigated. As a result of this study, a method enabling automated fuzzing testing for graphical interface applications on the Windows operating system. A software prototype was created to realize the proposed method and subjected to testing. 10.48612/jisp/fkpp-54ek-az1a 004.056 Automated fuzzing testing large language models graph representation https://jisp.spbstu.ru/article/2025.23.1/ pib_3_5-6.pdf

RAR RUS 20-29 0009-0004-1271-709X Peter the Great St. Petersburg Polytechnic University Ognev Roman ognev_ra@spbstu.ru 13103571000 0000-0002-0232-7248 Peter the Great St. Petersburg Polytechnic University Zegzhda Dmitry zegzhda_dp@spbstu.ru

Russia, 195251, St. Petersburg, Polytechnicheskaya str., 29

Generalized method for comparative analysis of fuzz testing tools A systematic analysis of fuzzing evaluation methodologies has been conducted. A minimal yet comprehensive metric set – branch coverage, unique crash count, and time-to-first-crash – has been identified. A normalized aggregate indicator is proposed that enables post-hoc comparison of different tools without reruns. 10.48612/jisp/rhtx-2n4n-dm6t 004.56 Fuzzing fuzzing metrics Large Language Models (LLM) vulnerability detection fuzz testing https://jisp.spbstu.ru/article/2025.23.2/ pib_3_5-6.pdf

RAR RUS 30-39 0000-0003-1345-1874 Peter the Great St. Petersburg Polytechnic University Pavlenko Evgeny pavlenko_eyu@spbstu.ru 0009-0004-3025-261X Peter the Great St. Petersburg Polytechnic University Ivanov Mikhail ivanov2.ms@edu.spbstu.ru Analysis of methods for detecting malicious software using large language models and eBPF technology A review of malware detection software tools using large language models and eBPF technology has been performed. For each tool under consideration, a brief description is provided, as well as its advantages and disadvantages. The results of a comparative analysis of the considered tools are presented, which make it possible to identify research areas in the field under consideration that require the most attention. 10.48612/jisp/bx59-dfz8-adbv 004.056 Information security malware Android LLM eBPF https://jisp.spbstu.ru/article/2025.23.3/ pib_3_5-6.pdf

RAR RUS 40-54 0000-0001-5757-381X Emperor Alexander I St. Petersburg State Transport University Gofman Maksim gofman@moonmail.ru 7006566675 0000-0002-6076-7241 Emperor Alexander I St. Petersburg State Transport University Kornienko Anatoliy kaa.pgups@yandex.ru

Russia, 190031, St. Petersburg, Moskovsky ave., 9

A method for predicting the correlation properties of marker bipolar sequences using digital watermarking by the patchwork method This article develops theoretical justifications and presents simulation results confirming the key influence of expanding bipolar sequences on the correlation properties of marker (expanded) bipolar sequences. Specifically, the article develops a method for predicting the values of the normalized autocorrelation function of an expanded (marker) bipolar sequence, and specifies the conditions for achieving high prediction accuracy. The article also proves the existence of a connection between the properties and characteristics of the mutual correlation function between the sought-after (embedded) marker sequence and distorted (by noise or attacks) sequence and the properties of the autocorrelation function of the embedded marker sequence. Finally, the article concludes with the formulation of indicators and criteria for the robustness of expanding and marker (extended) bipolar sequences. 10.48612/jisp/e4d3-afn7-x99z 519.876.5:519.6:004.357 Steganography digital watermarking bipolar sequences pseudo-random sequences information security copyright protection patchwork method autocorrelation function mutual correlation function https://jisp.spbstu.ru/article/2025.23.4/ pib_3_5-6.pdf

RAR RUS 55-68 0009-0005-6983-8017 National Research Nuclear University MEPhI Mandrov Aleksandr sashamandrovp@gmail.com 0009-0007-8532-3640 National Research Nuclear University MEPhI Kravchenko Nikita nik.kravchenko.2004@bk.ru 55229487100 0000-0002-4429-8799 National Research Nuclear University MEPhI (Moscow Engineering Physics Institute) Zhukov Igor i.zhukov@inbox.ru

Russia, 115409, Moscow, Kashirskoe shosse, 31

0009-0004-0370-1379 National Research Nuclear University MEPhI Balashova Ekaterina ekatherina04@gmail.com Building a semantic space of intentions using generative pre-trained models for solving the spam filtering task One of the key elements in solving spam message filtering is the text vectorization method. The article proposes a vectorization approach based on matching text to pairs of intentions. A list of intention pairs was identified and a synthetic dataset of textual utterances was generated. A neural network was designed and trained to determine the degree of belonging of each intention to the text expression at the model input. The developed method was tested on the spam message filtering task using logistic regression and the Enron dataset and SMS dataset 10.48612/jisp/rk44-9aab-nxha 004.81 Transformers DistilBert PCA neural networks synthetic data https://jisp.spbstu.ru/article/2025.23.5/ pib_3_5-6.pdf

RAR RUS 69-80 0000-0003-2273-725X Rostov State University of Economics Lapsar Aleksey lapsar1958@mail.ru

Russia, 344000, Rostov-on-Don, Bolshaya Sadovaya str., 69

0009-0007-7981-1146 Rostov State University of Economics Chizhevsky Maxim ch.inc@yandex.ru

Russia, 344000, Rostov-on-Don, Bolshaya Sadovaya str., 69

0000-0002-5583-4972 Rostov State University of Economics Serpeninov Oleg serpeninov53@mail.ru

Russia, 344000, Rostov-on-Don, Bolshaya Sadovaya str., 69

Optimization of computer incident investigation algorithm in SIEM systems In the context of increasing frequency and complexity of cyberattacks, effective incident investigation has become a priority task in ensuring organizational information security. One of the key challenges in using SIEM systems for investigating computer incidents is the lack of formalized algorithmic approaches to the processing and analysis of security events. To address this issue, an algorithm has been developed to optimize the actions of specialists when analyzing suspicious activity in information systems. The algorithm covers the key stages of investigation – from event verification to the analysis of potential intruder actions. The results of the study demonstrate that formalizing investigation processes contributes to more effective incident response and reduces the time required for their resolution. 10.48612/jisp/rmzt-68hn-ung8 004.056.5 Incident investigation algorithm information security SIEM systems event analysis incident response cybersecurity https://jisp.spbstu.ru/article/2025.23.6/ pib_3_5-6.pdf

RAR RUS 81-90 0000-0002-9732-0099 Peter the Great St. Petersburg Polytechnic University Kalinin Maxim max@ibks.spbstu.ru

Russia, 195251, St. Petersburg, Polytechnicheskaya str., 29

Peter the Great St. Petersburg Polytechnic University Vasiliev Oleg vasiliev_os@spbstu.ru 0000-0002-2264-7513 Peter the Great St. Petersburg Polytechnic University Krundyshev Vasiliy krundyshev_vm@spbstu.ru Ensuring byzantine fault tolerance of a distributed ledgers in a smart city based on a trust model The paper examines the problem of ensuring Byzantine fault tolerance in smart city distributed ledger systems. To improve the security of distributed ledgers, it is proposed to use the hashgraph distributed consensus protocol, in which events are organized as a directed acyclic graph, and nodes exchange “gossip about gossip”. The traditional hashgraph protocol has been supplemented with the developed trust model that reduces the impact of malicious devices when reaching consensus in digital infrastructures of a smart city. 10.48612/jisp/39ak-3h78-b99n 004.056 Byzantine fault tolerance consensus trust model directed acyclic graph distributed ledger node reputation smart city DAG hashgraph https://jisp.spbstu.ru/article/2025.23.7/ pib_3_5-6.pdf

RAR RUS 91-99 Peter the Great St. Petersburg Polytechnic University Zavadskii Evgeniy zavadskij_ev@spbstu.ru Cybersecurity of the functional network of critical infrastructure facilities using the strategy of optimal placement of Honeypot nodes The article considers the problem of dynamic management of a protection system based on Honeypot technology. To solve this problem, a method is proposed to reduce the requirements for computing resources and a function for assessing the criticality and risk of compromising nodes of the functional network infrastructure, the use of which allows taking into account the peculiarities of the implementation of technological processes and the dynamics of attack development. 10.48612/jisp/tp88-3x44-g92n 004.94 Network security Honeypot multicriteria optimization dynamic network cyberattack graph model https://jisp.spbstu.ru/article/2025.23.8/ pib_3_5-6.pdf

RAR RUS 100-109 0000-0001-7199-6384 Special Technology Center LLC Kurakin Alexander akurakin@stc-spb.ru Method of dynamic distribution of access rights in an autonomous group of robotic systems based on genetic algorithms A method for delimiting access rights in a group of heterogeneous robotic systems is proposed based on the concept of a virtual squad, considering their information interaction, based on a genetic algorithm. The essence of the method is that when using a role model for delimiting the rights of access of a subject to an object – the goal of a work task, the model of delimiting access rights is updated to exclude a collision. To change the parameters of the model, the roles are presented as a set of binary data that form a chromosome of the initial population, which changes based on a genetic algorithm when the current situation changes during the work task. Modeling is carried out and the results of numerical studies are presented in comparison with the solution of the problem by the Hungarian algorithm. 10.48612/jisp/v4hu-91h9-5m6m 004.021 Grouping of robotic systems dynamic distribution method algorithm mission https://jisp.spbstu.ru/article/2025.23.9/ pib_3_5-6.pdf

RAR RUS 110-120 0009-0005-6662-5606 Saint Petersburg State University of Aerospace Instrumentation Velichko Ivan wwr0ngn4m3@gmail.com

Russia, 190000, St. Petersburg, Bolshaya Morskaya str., 67, liter A

0000-0002-0924-6221 Saint Petersburg State University of Aerospace Instrumentation Bezzateev Sergey sergey.bezzateev@gmail.com

Russia, 190000, St. Petersburg, Bolshaya Morskaya str., 67, liter A

From exploitation to protection: analysis of methods for defending against attacks on LLMS Modern large language models demonstrate impressive capabilities but remain vulnerable to attacks that can manipulate their behavior, extract confidential data, or bypass built-in restrictions. This paper focuses on methods for protecting language models from prompt injection attacks, which allow adversaries to exploit the system for malicious purposes. Various defense strategies are examined and analyzed, including query filtering, context isolation, training on perturbed data, and other approaches. A comparative analysis of the effectiveness of defense mechanisms is conducted, highlighting their limitations and identifying future directions for enhancing the security of language models. 10.48612/jisp/2nkk-unee-uuzv 004.056.52 Large language models artificial intelligence adversarial attacks defense methods model output manipulation https://jisp.spbstu.ru/article/2025.23.10/ pib_3_5-6.pdf

RAR RUS 121-146 MIREA – Russian Technological University Matuhina Ekaterina matyuhina@mirea.ru 0000-0002-7231-5728 MIREA – Russian Technological University Spirin Andrey spirin_aa@mirea.ru 0009-0004-0833-5574 MIREA – Russian Technological University Ikonnikov Aleksandr alx.ikona@gmail.com Cluster analysis of vector representations of malicious query language models: comparison of methods of obtaining embeddings based on character N-grams, individual words a nd whole sentences This study presents a comparative analysis of tokenization strategies and text vectorization methods for detecting harmful jailbreak prompts submitted to large language models. Using a dataset of both benign and malicious queries, three approaches were evaluated: aggregated embeddings of character-level N-grams, aggregated word-level embeddings, and semantic representations of entire prompts. The results show that token-based methods achieve a high recall of malicious prompts by capturing repetitive local patterns, though often at the cost of increased false positives. In contrast, semantic embeddings of full prompts provide high precision in detecting threats, but may overlook obfuscated or rare attacks. A key finding of this work is that vector representations demonstrate clear cluster separation between benign and harmful prompts, making it possible to apply lightweight classification algorithms for effective filtering, especially in systems with limited computational resources. The study also supports a two-stage protection framework, where clustering is used as a preliminary filter, and only suspicious inputs proceed to deeper analysis. In some configurations, the approach successfully identified up to 96 % of jailbreak prompts, confirming its practical relevance for integration into large language model access pipelines. 10.48612/jisp/5ua4-umte-db1n 004.056.5, 004.85:004.89, 519.237.8 Language models text tokenization semantic embeddings clustering of vector representations filtering of malicious requests bypassing model restrictions https://jisp.spbstu.ru/article/2025.23.11/ pib_3_5-6.pdf

RAR RUS 147-164 0000-0002-3830-1840 Peter the Great St. Petersburg Polytechnic University Yugai Pavel yugaj_pe@spbstu.ru Research of adversarial attacks on classical machine learning models in the context of network threat detection This study presents an investigation of adversarial attacks on classical machine learning algorithms within the context of network threat detection. It offers an overview of the machine learning models employed for various tasks in the realm of computer network security. A formal description of the threat model is provided, along with a classification of adversarial attacks. The classification of network traffic within the WEB-IDS23 dataset is carried out using classical machine learning models, including k-nearest neighbors, random forest, and support vector machine. Implemented adversarial attacks include the Fast Gradient Sign Method, Projected Gradient Descent, the Carlini and Wagner attack, and DeepFool, applied to these machine learning algorithms. An analysis of the impact of the deployed adversarial attacks on the aforementioned classical machine learning algorithms is conducted. 10.48612/jisp/xerr-dfhh-2zak 004.89 Adversarial attacks machine learning network threats Fast Gradient Sign Method Projected Gradient Descent Carlini and Wagner attack k-nearest neighbors support vector machine random forest https://jisp.spbstu.ru/article/2025.23.12/ pib_3_5-6.pdf

RAR RUS 165-179 0000-0001-9659-1244 Peter the Great St. Petersburg Polytechnic University Poltavtseva Maria potavtseva@ibks.spbstu.ru Optimization of data obfuscation in big data processing and storage systems The paper is devoted to the task of reducing the attack surface from an internal attacker in heterogeneous big data processing and storage systems by choosing the optimal method of data obfuscation based on anonymization (depersonalization) technologies. The paper analyzes 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 method of data obfuscation and an algorithm for solving it over various types of datasets are proposed, taking into account evaluation criteria specific to each class of methods. The implementation of a software prototype to support decision-making and the choice of the optimal method for solving practical problems is described, experimental approbation and analysis of its results are carried out. 10.48612/jisp/693e-m24n-96zh 004.056 Big data security big data management systems data privacy data obfuscation data anonymization https://jisp.spbstu.ru/article/2025.23.13/ pib_3_5-6.pdf

RAR RUS 180-191 0000-0002-5606-7509 MIREA – Russian Technological University Mironkin Vladimir mironkin.v@mail.ru National Research University Higher School of Economics Yurasov Nikita n.yurasov@yahoo.com On the question of the applicability of the statistics of the measure “chi-square” in the problem of determining the parameters of block encryption algorithms The article presents the results of the practical application of statistical methods for analyzing block encryption algorithms operating in the ECB encryption mode. In particular, in cases of incorrect use of plaintext encoding algorithms, it is established that it is possible to determine the block size of encryption algorithm based on the analysis of chi-square statistics calculated on the available ciphertext volume. 10.48612/jisp/3p1p-bx2v-prru 004.056.55 Block encryption algorithm encryption mode ECB mode statistics criterion chi-squared distribution https://jisp.spbstu.ru/article/2025.23.14/ pib_3_5-6.pdf

RAR RUS 192-212 0009-0008-9694-2348 Russian Presidential Academy of National Economy and Public Administration (RANEPA) Biryukova Anastasiya nastyabir1010@gmail.com 0000-0003-1300-2470 Mozhaysky Military Space Academy Biryukov Denis Biryukov.D.N@yandex.ru

Russia, 197198, St. Petersburg, Zhdanovskaya str., 13

0000-0001-9665-0128 Peter the Great St. Petersburg Polytechnic University Suprun Alexander afs54@inbox.ru

Russia, 195251, St. Petersburg, Polytechnicheskaya str., 29

Global practices in regulating and implementing generative artificial intelligence in higher education This article analyzes international recommendations for the application of generative artificial intelligence (AI) in higher education over the past five years. It identifies key trends, ethical challenges, and strategies for integrating AI technologies into academic practices. The study focuses on the tensions between AI’s innovative potential (e. g., personalized learning, automation of routine tasks) and associated risks (e. g., academic dishonesty, digital inequality). Regulatory initiatives, such as the EU AI Act, China’s AI standards, and developers’ ethical declarations, are examined alongside successful implementation practices, including MIT’s adaptive learning platforms and SberUniversity’s AI-driven digital assistants. Key findings emphasize: the need to balance technological progress with ethical norms, including mandatory AI-generated content labeling and the promotion of AI literacy; the importance of global standards to overcome legal fragmentation and bridge the digital divide. Recommendations for universities: phased AI integration, infrastructure investments, and staff training programs. The study contributes to shaping strategies for adapting higher education to the era of generative AI, highlighting universities’ role as drivers of responsible technology adoption. The findings are relevant for university administrators, policymakers, and EdTech developers. 10.48612/jisp/hz8x-hdxv-b83d 004.8:378.147 Generative artificial intelligence higher education cybersecurity academic integrity personalized learning AI ethics https://jisp.spbstu.ru/article/2025.23.15/ pib_3_5-6.pdf