Sun’iy Intellekt va Raqamli Iqtisodiyot (SIRI 2025)

TECHNOLOGY-ASSISTED LANGUAGE LEARNING IN SECOND LANGUAGE ACQUISITION Raxmonova Guzal 706 THE EVOLUTION OF TECHNOLOGY AND ITS INTRODUCTION TO THE SPHERE OF EDUCATION Rahmonova Gulira’no 710 LINGUISTIC AND SOCIO-HUMANITARIAN ISSUES OF TRAINING PERSONNEL AND ORGANIZING EDUCATION BASED ON ARTIFICIAL INTELLIGENCE AND INNOVATIVE APPROACHES Shukurilloyeva Lazzatoy 714 BOLALARNI FIKRLASHINI RIVOJLANTIRISHGA KO‘MAKLASHADIGAN MOBIL O‘YIN DASTURLARINING AHAMIYATI Bahronova Nilufar 717 TARJIMA JARAYONLARIDA QO‘LLANILADIGAN SUN’IY INTELLEKTLAR: ULARNING AFZALLIKLARI VA KAMCHILIKLARI, TARJIMA JARAYONIDA TA’SIRI Zoirova Azizabonu 721 YANGI O‘ZBEKISTONDA TA’LIMGA E’TIBOR VA OLIB BORILAYOTGAN ISLOHOTLAR KO‘LAMI Qilichova Odina 726 AI AS AN ASSISTANT FOR TRANSLATORS Aldanova Guljakhon 729 HOW TRAVELING ABROAD IMPROVES LANGUAGE LEARNING Qodirova Farangiz 733 MAKTABGACHA YOSHDAGI BOLALARNI TARBIYA QILISHDAGI OMILLARNING AHAMIYATI XUSUSIDA Madraximova Irodaxon 737

XITOY TILINI CHET TILI SIFATIDA O‘RGANISH VA O‘RGATISH: ILMIY YONDASHUVLAR VA METODOLOGIYALAR Mardonova Niginabonu 741 FUNCTIONAL-SEMANTIC FEATURES OF ECOTERMS AND THEIR LEXICOGRAPHICAL INTERPRETATION Arifjanova Nilufar 748 OILADA TARBIYA ASOSLARINI TAHLIL QILISH, UNING JAMIYATDAGI AHAMIYATI XUSUSIDA Madraximova Irodaxon 752 HOW ONLINE LEARNING IS CHANGING TRADITIONAL CLASSROOMS Po‘latova Sevara 756 TIL O‘RGANISH JARAYONIDA SUN’IY INTELLEKTNING O‘RNI VA UNDAN UNUMLI FOYDALANISH Sunatillayeva E‘zoza 760 THE IMPACT OF ARTIFICIAL INTELLIGENCE ON THE FIELD OF TRANSLATION AND ITS FUTURE PROSPECTS Egamnazarova Rushana 764 ФАРЗАНД ТАРБИЯСИ: ТУҒИЛИШДАН 17 ЁШГАЧА БУЛГАН БОЛАЛАРНИНГ БОСҚИЧМА-БОСҚИЧ РИВОЖЛАНИШИ ХУСУСИДА Хаётхон Хакимова 770 IJTIMOIY PEDAGOGIKADA MEʼYORDAN CHEKINGAN OʻSMIRLAR BILAN OLIB BORILADIGAN ISH SHAKLLARI, METODLARI VA VOSITALARI Omina Foziljonova 775 HOW TO BOOST WRITING SKILLS IN LEARNING FOREIGN LANGUAGES Ilhomberdiyeva Husnora 778 XITOY IEROGLIFLARINING O‘ZIGA XOS XUSUSSIYATLARI Temirova Shaxloxon 781

IV-SHO‘BA: Sun’iy intellekt davrida matematik va fizik modellashtirishning o‘rni TURLI KO‘NDALANG KESIM YUZASIGA EGA SUVLI KOLLEKTOR QISMLARI UCHUN DIFFUZIYA MASALASI Alisher Usmonov, Shaxlo Xabibullayeva, Hasan Xolmatov 785 МАТЕМАТИЧЕСКОЕ МОДЕЛИРОВАНИЕ В НЕКОТОРЫХ ЭКОНОМИЧЕСКИХ ЗАДАЧАХ Кувончбек Асадов 790 TALABALARGA GIPOTSIKLOIDALAR GRAFIKLARINI KOMPYUTER GRAFIKASIDAN FOYDALANIB CHIZISHNI O‘RGATISH HAQIDA Feruza Raximova, Munira Payziyeva, Nilufar Pardayeva, Zarina Saidova, Quvondiq Begimbetov 794 SUN’IY INTELLEKT YONDASHUVLARI ASOSIDA YUPQA PLASTINALARNING TERMO-MAGNIT ELASTIK DEFORMATSIYALANISH JARAYONLARINI MATEMATIK MODELLASHTIRISH Mirzaaxmedov Muxammadbobur, Ovxunov Iqboljon 799 NANOFIZIKADA SUN'IY INTELLEKTNING QO‘LLANILISHI HAMDA TIBBIYOT SOHALARDA NAMOYON BO‘LISHI U. M. Naziraliyeva 802 TELEKOMMUNIKATSIYA ELEKTR TA’MINOTI TIZIMINI MATEMATIK MODELLASHTIRISH Fazliddin Qodirov 807 FIZIKA SOHASIDA AUGMENTED REALITY VA VIRTUAL REALITY TEXNOLOGIYALARIDAN FOYDALANISH U. M. Naziraliyeva 812 КОМПЬЮТЕРНАЯ МОДЕЛЬ РАСПРОСТРАНЕНИЯ ПРИМЕСИ В АТМОСФЕРЕ Малик Убайдуллаев 814

SUN'IY INTELLEKT YORDAMI BILAN KATTA O‘LCHAMLI OPTIMIZATSIYA VA KOMBINATORIK MUAMMOLARNI YECHISH 819 Abduqaxorova Mashxurabonu GAMMA FUNKSIYASI YORDAMIDA KASR SONLAR FAKTORIALINI HISOBLASHNI MODELLASHTIRISH VA DASTURINI YARATISH HAQIDA Quvondiq Begimbetov, Sarvar Safarboyev, Diyorbek Mamasoliyev, Dilafruz Iskandarova 824 РОЛЬ МОДЕЛИРОВАНИЯ ФИЗИКИ В ЭПОХУ ИСКУССТВЕННОГО ИНТЕЛЛЕКТА 829 Xasanova Lolaxon, Rashidov Ilyas ЭФФЕКТИВНЫЕ АЛГОРИТМЫ БЫСТРОГО УМНОЖЕНИЯ ЧИСЕЛ: АНАЛИЗ МЕТОДОВ И ОПТИМИЗАЦИОННЫХ ПОДХОДОВ Surojiddin Pardaboyev, Tojixonov Javohir, Abdulov Ibroxim 833 V-SHO‘BA: Raqamli texnologiyalar davrida matematik va fizik modellashtirishning o‘rni THE ROLE OF PROGRAMMING LANGUAGES IN SOFTWARE DEVELOPMENT Abdulakhatov Muzaffarjon 838 AVTOMATLASHTIRILGAN TASHXIS VA DAVOLASH JARAYONLARINI LOYIHALASHNING EHTIMOLIY MODELINI YARATISH Odilbek Asqaraliyev, Muslimjon Tolibjanov 843 HUDUDNING IJTIMOIY-IQTISODIY RIVOJLANISHINI “YASHIL” IQTISODIYOT ASOSIDA SHAKLLANTIRISHNING ME’YORIY-HUQUQIY ASOSLARI TAHLILI (SURXONDARYO VILOYATI MISOLIDA) Shomirzayev Abdug‘affor 848

RAQAMLI KELAJAK: SUN'IY INTELLEKT VA TEXNOLOGIYALARNING AFZALLIKLARI VA KAMCHILIKLARI Mirxamidova Madinabonu, Ro‘ziboyev Diyorbek, Narmanov Otabek 855 ELEKTRON DARSLIKLARNI YARATISHDA AUTOPLAY MEDIA STUDIO DASTURINI QO‘LLASH Surxonov Jamshid 859 IMLOVIY XATOLARNI TEKSHIRISHGA MO‘LJALLANGAN INTERAKTIV DASTUR IMKONIYATLARI RAXMATOVA KUMUSHOY THE EFFECT OF TECHNOLOGICAL INNOVATIONS ON GREEN DIGITAL ECONOMIES AND DEVELOPMENT STRATEGIES Mashkhura Azizjon 866 ONLINE KUTUBXONA MOBIL ILOVASI IMKONIYATLARI VA TAVSIFI Abdurashidova Mohinur 871 XUSUSIY TIBBIYOT MARKAZI MIJOZLARI VA XODIMLARI UCHUN AXBOROT TIZIMINI YARATISH Xaitova Odina 875 FUNKSIONAL DIAGNOSTIKA BO‘LIMIDAGI ELEKTROKARDIOGRMMA TEXNIKASINI SOZLASH, TEXNIK XIZMAT KO‘RSATISH VA DASTURIY TA'MINOTI YAXSHI ISHLASHINI TEKSHIRISHDA SUN'IY INTELEKTDAN SAMARALI FOYDALANISH BOSQICHLARI Mirzakulova Gulbahor 879

I – SHO‘BA Sun’iy intellekt va raqamli texnologiyalarning muammolari va yechimlari

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 Enhancing Multimodal Image Retrieval for Uzbek language through CLIP Architecture Adaptation Mukhriddin Mukhiddinov1,*, Utkir Khamdamov1, Oybek Djuraev1, Jasurbek Djuraev2 1Department of Communication and Digital Technologies, University of Management and Future Technologies, Tashkent, Uzbekistan. mukhiddinov@umft.uz 2Department of Software Engineering, Tashkent University of Information Technologies named after Muhammad al-Khwarizmi, Tashkent, Uzbekistan. jasrbek.djuraev.djn0926@gmail.com Abstract: This study investigates the development of multimodal vision-language models for image retrieval in low-resource languages, with a specific focus on Uzbek. Existing vision-language models are predominantly designed for high-resource languages, and their fine-tuning typically requires substantial computational resources. To address the limitations associated with image-text retrieval in low-resource linguistic settings, this research adopts the CLIP model architecture and incorporates a range of strategies aimed at optimizing the trade-off between computational efficiency and model performance. These strategies include the generation of synthetic data via machine translation, image augmentation techniques, and targeted training of attention mechanisms in transformerbased models using domain-specific datasets. The proposed framework integrates Multilingual BERT as the text encoder alongside various image encoders, including ResNet50, EfficientNet-B0, Vision Transformer (ViT), and Tiny Swin Transformer. Experimental results indicate that EfficientNet-B0 and Tiny Swin Transformer yield superior performance on benchmark datasets such as COCO, Flickr30k, and Flickr8k. Key words: visual question answering, blind and visually impaired; deep learning; lowresource languages; CLIP Architecture. Introduction The exponential growth of digital content—encompassing text, images, and videos— has posed significant challenges for traditional information retrieval systems, which are primarily designed for textual input. Keyword-based search methods often return overly broad results that fail to accurately reflect user intent or the complexity of multimedia content. An ideal retrieval system should facilitate access to relevant information irrespective of the user’s native language or preferred input modality. Multimodal image retrieval addresses this need by enabling queries through various forms, including text, images, spoken language, or a combination thereof. This multimodal approach significantly enhances search capabilities and broadens access to information across diverse linguistic and interaction preferences. For example, image-to-image retrieval allows users to photograph a building to identify its architectural style or submit a sketch to search for similar clothing items online. Such applications demonstrate the potential of multimodal search in delivering more intuitive and effective user experiences. However, a major obstacle persists: most existing 27

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot multimodal retrieval systems rely on large-scale, resource-intensive models trained on massive datasets, predominantly in high-resource languages. This creates a substantial barrier for low-resource languages, which lack the extensive annotated data necessary for training such models. This research seeks to bridge this gap by improving the memory and computational efficiency of multimodal retrieval systems while harnessing the wealth of image data available across languages. The goal is to make these advanced technologies more inclusive and accessible. Notably, current multimodal systems often exhibit diminished scalability and performance in low-resource settings, particularly when dealing with highdimensional, large-scale data. Furthermore, disparities in the availability of different data types pose additional difficulties for training effective multimodal models in underrepresented languages with limited digital content. The rationale for advancing multimodal capabilities in low-resource (LR) languages is both timely and well-founded. Firstly, multimodal processing aligns more closely with natural human communication, which inherently involves the integration of multiple sensory inputs. Secondly, visual and auditory information can offer essential contextual cues that help mitigate the challenges posed by limited textual data. Thirdly, many LR languages are predominantly oral in nature, rendering multimodal strategies particularly suitable for their digital representation, preservation, and computational processing. Despite these advantages, the development of multimodal systems for LR languages is confronted with several notable challenges. These include: (1) the limited availability of high-quality multimodal datasets specific to these languages; (2) the absence of standardized evaluation metrics and benchmarks; (3) the high computational demands associated with training large-scale models under constrained resource conditions; and (4) the added complexity of addressing diverse scripts, dialectal variations, and culturally specific contexts. Additionally, technical issues such as catastrophic forgetting during the adaptation of pre-trained models to new linguistic environments, and the difficulty of maintaining robust performance across multiple modalities, further complicate the development of effective systems for LR languages. The principal aim of this study is to design a multimodal vision-language retrieval system tailored for the Uzbek language, with an emphasis on achieving a balance between computational efficiency and scalability within low-resource environments. A key aspect of this research involves investigating strategies for optimizing model performance across various domains, particularly in the context of limited data availability. The objective is to develop a robust system that not only performs effectively for Uzbek but also serves as a scalable framework adaptable to other low-resource languages such as Kazakh and Azerbaijani. Through this approach, the study aspires to extend the applicability of advanced AI technologies to linguistically diverse and resource-constrained settings, thereby democratizing access to computationally intensive solutions for a wider global community. Related works Recent progress in image retrieval has been largely driven by the integration of deep learning methodologies, marking a transition from traditional keyword-based approaches to more sophisticated context-aware techniques. Early innovations, such as the color indexing 28

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 method proposed by Swain and Ballard, represented a foundational shift by utilizing inherent image features rather than relying solely on externally assigned keywords. The introduction of Convolutional Neural Networks (CNNs) further advanced the field by enabling the extraction of high-level semantic features from images, thereby enhancing the accuracy and relevance of retrieval systems. The development of multimodal learning, as demonstrated in the work [1], introduced the joint modeling of textual and visual data to learn unified deep representations. Subsequent models, including AlexNet and more recent transformer-based architectures, have refined the processes of feature extraction and matching through the use of visual vocabularies. These advancements have significantly improved the efficiency of image retrieval systems and facilitated real-time processing capabilities, as evidenced in the studies by [2] and [3]. Furthermore, the distribution of research attention across languages in the field of multimodal learning remains markedly imbalanced. As illustrated in Figure 1, languages such as Hindi (26 publications), Arabic (18 publications), and Bengali (16 publications) have received comparatively extensive attention, while approximately 40 other languages are represented by only a single study each. This disparity underscores the pressing need to expand research efforts toward underrepresented languages within multimodal frameworks. Recent surveys have examined various dimensions of multimodal language modeling. For instance, [4] offer a comprehensive review of large multimodal model (LMM) architectures, training methodologies, and practical applications, whereas [5] concentrate on pre-training strategies and model design. Nonetheless, these surveys predominantly address high-resource languages and fall short of exploring the specific challenges and methodological adaptations required for low-resource languages in multimodal settings. Figure 1. The distribution of 106 publications across 75 low-resource languages reveals a pronounced imbalance in research attention. While a small subset of languages has garnered relatively greater focus, the majority remain significantly underrepresented within the scope of multimodal learning. This data underscores the need for more inclusive research efforts targeting a broader range of low-resource languages. 29

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot Figure 2 presents a structured taxonomy of large multimodal models (LMMs) developed for low-resource (LR) languages, categorized into six primary research directions. This classification captures both the current landscape of the field and the key methodological strategies employed to address the unique challenges faced by underrepresented languages. The first two categories emphasize the development of highquality resources. The initial category pertains to the construction of multimodal datasets, either through original data collection or by extending existing corpora. The second category focuses on synthetic data generation, which expands limited resources through techniques such as back-translation and image-based data synthesis. Figure 2. A high-level taxonomy of large multimodal models (LMMs) for low-resource languages is presented, comprising six primary categories (highlighted within greenshaded boxes). Each main category is further divided into subcategories, illustrated through selected representative studies. Building upon these foundational efforts, the third category explores multimodal fusion strategies, including early fusion, late fusion, and more advanced hybrid techniques for integrating information across modalities. The fourth category highlights visual enhancement methods, such as image-guided translation and visual disambiguation, which leverage visual context to improve translation accuracy and resolve linguistic ambiguities. The fifth category transitions from unimodal approaches to cross-modal transfer learning, facilitating knowledge transfer across both modalities and languages to improve model generalization in low-resource settings. The final category encompasses architectural innovations specifically designed to optimize multimodal learning for LR languages. The remainder of this article is organized according to the proposed taxonomy illustrated in Figure 2. Research methodology The system is structured around two primary components: an image encoder and a text encoder, which collaboratively process and align visual and textual data. The image encoder utilizes a pretrained network, such as ResNet50, which is modified to exclude the final classification layer, focusing instead on extracting feature representations from the images. The text encoder is based on a pretrained BERT model, designed to handle multiple 30

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 languages, thereby enabling it to process text in diverse linguistic contexts (BertModel from Hugging Face). To facilitate the alignment between image and text representations, both encoder outputs are passed through a projection head. This component is responsible for reducing the dimensionality of the outputs, ensuring they are of comparable size. The comparison of image and text features is then conducted using contrastive loss, a technique that not only enables the model to relate corresponding images and texts but also guides it in learning which visual and textual elements are contextually connected. This research leverages a combination of established and newly curated datasets, which have been augmented with Uzbek translations to support more effective training in a low-resource linguistic context. A key resource utilized is the MS COCO dataset, renowned for its comprehensive image annotations and object segmentation, which facilitate advanced image understanding and recognition in naturalistic environments as shown in Figure 3 [6]. Figure 3. COCO Dataset: Example images and corresponding captions from the Microsoft COCO Caption dataset [6]. Another significant dataset is Flickr30k, which comprises approximately 31,000 images sourced from Flickr, each accompanied by five descriptive captions. These annotations offer rich contextual details regarding scenes, objects, and actions, thereby 31

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot enhancing the model’s ability to interpret and generate semantically meaningful textual representations of visual content [7]. Additionally, the ROCO (Radiology Objects in Context) dataset provides specialized medical imaging data, including radiological scans paired with corresponding textual descriptions. This dataset enables the model to extend its capabilities to the medical domain by learning to interpret and correlate visual medical content with textual information [8]. The architecture illustrated in Figure 5 demonstrates an effective approach for processing both visual and textual modalities, particularly enhancing performance in visionlanguage tasks under low-resource conditions by minimizing data and computational requirements. By combining image features and textual embeddings within a unified framework, and leveraging lightweight models, the system reduces the need for extensive fine-tuning. Its encoding efficiency is further improved through techniques such as image augmentation and the incorporation of domain-specific textual corpora, which contribute to refining the attention mechanisms within a multilingual BERT framework. The visual encoder employs pre-trained convolutional neural networks (CNNs), with the final classification layer replaced by an identity function to retain rich visual representations. A range of CNN-based and transformer-based architectures—including ResNet-50, EfficientNet-B0, Vision Transformer (ViT), Swin Transformer, and ConvNeXt—were explored for their capacity in multimodal feature extraction. These models were assessed based on metrics such as top-1 and top-5 accuracy on the ImageNet-1K dataset, parameter count, GFLOPS, and model size, thereby informing the selection of the most suitable architecture according to specific application needs. Figure 5. Projection Head: A framework for calculating compatibility scores between image-text pairs, adapted from [3]. We utilized the multilingual BERT model [9] as the text encoder, which is well-suited for low-resource languages due to its tokenization capabilities spanning over 104 32

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 languages. This makes it particularly effective for applications demanding broad multilingual support (refer to Figure 4). Figure 4. Training strategy of Multilingual BERT utilizing masked language modeling, adapted from [9]. Experiment Design This section presents the evaluation methodology employed to assess the performance of the developed vision-language models. The evaluation was designed to examine the models' ability to generalize across various datasets, providing insights into their performance on both previously seen and entirely novel data. In-domain testing involved evaluating the models on the datasets used during training to assess their effectiveness in familiar environments. Conversely, out-of-domain testing evaluated model performance on unseen datasets, offering a measure of their generalizability and robustness across diverse data distributions. Our experimental setup incorporated a range of visual encoders integrated with a base Multilingual BERT model for text processing, with a focus on the Uzbek language. The visual architectures included ResNet-50, EfficientNet-B0, Vision Transformer (ViT), and Tiny Swin Transformer. Datasets used in the experiments comprised MSCOCO, Flickr8k, and Flickr30k. Additionally, synthetic datasets were created through machine translation and data augmentation to replicate low-resource language scenarios. Evaluation metrics included Mean Average Precision (MAP), Mean Average Recall (MAR), Mean Average F1-Score (MAF1), and Top-K Accuracy (Top-1, Top-5, and Top-10), providing a comprehensive view of model performance. This section systematically outlines the experimental variables that were altered across different trials, along with the rationale for each modification aimed at analyzing their influence on model performance. These variations provided valuable insights into the 33

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot optimization and scalability of the models across diverse scenarios, particularly in lowresource and cross-domain contexts. Model Configurations: A range of configurations for both text and image encoders were explored, involving changes in the number of layers, activation functions, and the depth of fine-tuning. The objective was to identify optimal combinations that achieve a balance between performance and computational efficiency. The experiments included comparative analyses between the standard BERT architecture and its multilingual variant, as well as evaluations of various image encoder backbones—such as ResNet-50 and alternative architectures. These investigations aimed to understand the trade-offs between accuracy and computational cost across different configurations [10]. Conclusions This study presents the development of a multimodal vision-language framework for image retrieval in low-resource language settings, with a specific emphasis on the Uzbek language. By leveraging synthetic datasets—generated through machine translation from English to Uzbek and enhanced with data augmentation techniques—the proposed approach achieved improved retrieval performance compared to conventional models. Future research directions include the following: • Extension to Other Low-Resource Languages and Advanced Data Augmentation: Broadening the application of this framework to additional low-resource languages, while integrating more sophisticated data augmentation strategies and incorporating human validation of synthetic data, may further improve model accuracy and generalization. • Cross-Domain Evaluation of Model Architectures: Conducting evaluations across a variety of model architectures on translated versions of the ROCO dataset—particularly in specialized domains such as medical imaging—could yield deeper insights into model robustness and adaptability. • Image Captioning and Energy-Efficient Deployment: Investigating the reverse task of image captioning and optimizing the models for energy efficiency would facilitate their use in resource-constrained environments, such as deployment on mobile and embedded devices. • Interactive Feedback Mechanisms: Implementing user-centered feedback systems to iteratively refine model outputs may enhance both retrieval precision and user experience. Continued exploration in these areas will support the broader accessibility and realworld applicability of AI technologies, particularly in underserved linguistic and technological contexts. The open dissemination of configurations and the encouragement of collaborative development remain integral to advancing this field. References 1. Nitish Srivastava and Ruslan Salakhutdinov. 2012. Multimodal learning with deep boltzmann machines. Advances in neural information processing systems. 2. Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. 2016. Learning visual features from large weakly supervised data. European conference on computer vision. 34

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 3. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. Preprint, arXiv:2103.00020. 4. S. Yin, C. Fu, S. Zhao, et al. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024. 5. X. Wang, G. Chen, G. Qian, et al. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20:447– 482, 2023. 6. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer. 7. Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In Transactions of the Association for Computational Linguistics, volume 2, pages 67–78. 8. Oskar Pelka, Sven Koitka, Julius Rückert, Felix Nensa, and Christoph M Friedrich. 2018. Radiology objects in context (roco): A multimodal image dataset. In Intravascular Imaging and Computer Assisted Stenting, and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, volume 11043 of Lecture Notes in Computer Science, pages 180–189, Cham. Springer. 9. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019a. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics. 10. A. Asgarov and S. Rustamov. LowCLIP: Adapting the CLIP Model Architecture for Low-Resource Languages in Multimodal Image Retrieval Task. arXiv:2408.13909, 2024. 35

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot AXBOROT XAVFSIZLIGINI TA’MINLASHDA MAZMUNNI TEKSHIRISH ALGORITMLARI Baxtiyor Akmuradov Kommunikatsiya va raqamli texnologiyalar kafedrasi, “University of Management and Future Technologies”, Toshkent, O‘zbekiston. b.u.akmuradov@gmail.com Annotatsiya: Maqolada telekommunikatsiya tizimlarida axborot xavfsizligini ta’minlashda Data Loss Prevention (DLP) texnologiyalarining o‘rni va ularning asosiy yondashuvlari - Policy-based DLP va Content Inspection tahlil qilinadi. DLP algoritmlarining Deep Packet Inspection, Pattern Recognition, Heuristic Analysis, NLP va File Fingerprinting kabi zamonaviy usullari asosida ma’lumotlarning sizib chiqishi va o‘g‘irlanishining oldini olish mexanizmlari bayon etiladi. Shuningdek, Symantec DLP modullari misolida himoya arxitekturasi va modullararo integratsiya prinsiplari ko‘rib chiqiladi. Kalit so‘zlar: Axborot xavfsizligi, DLP, ma’lumotlar himoyasi, Content Inspection, Policy-based DLP, DPI, Pattern Recognition, Symantec DLP, tarmoq xavfsizligi. Kirish Hozirgi kunda axborot texnologiyalarining rivojlanishi bilan birga telekommunikatsiya tizimlarining axborot xavfsizligini ta’minlash masalasi ham dolzarb muammolardan biriga aylanib bormoqda. Telekommunikatsiya infratuzilmalari orqali har kuni katta hajmdagi ma’lumotlar uzatiladi, bu esa kiberxavfsizlik tahdidlari ehtimolini oshiradi. Raqamli axborot resurslarining himoyasizligi natijasida muhim ma’lumotlarning ruxsatsiz foydalanilishi, o‘g‘irlanishi yoki buzilishi xavfi mavjud bo‘lib, bu tashkilotlar, korxonalar va davlat idoralari uchun katta yo‘qotishlarga olib kelishi mumkin. Shu sababli, telekommunikatsiya tizimlarining xavfsizlik darajasini oshirish uchun zamonaviy himoya usullarini joriy etish va samaradorligini oshirish dolzarb masala hisoblanadi va bu borada davlat darajasidagi muhim masalalardan biri sifatida qaraladi[1]. Data Loss Prevention - bu korxonalar va tashkilotlarda ma’lumotlarning yo‘qolishi, o‘g‘irlanishi yoki ruxsatsiz foydalanilishini oldini olish uchun qo‘llaniladigan texnologiyalar, jarayonlar va siyosatlar majmuasi. DLP tizimlari tashkilotning maxfiy ma’lumotlarini, intellektual mulklarini va shaxsiy ma’lumotlarini himoya qilishga qaratilgan. Data Loss Prevention ya’ni yo‘qotishlarning oldini olish algoritmlarini shartli ravishda quyidagi ikki guruhga ajratish mumkin: - Policy-based DLP: Maxfiy ma’lumotlarga kirish huquqlarini boshqarish va xususiy ma’lumotlarning yo‘qolishining oldini olish. Policy-based DLP (Siyosatga asoslangan DLP) bu tashkilotda ma’lumotlarni himoya qilish siyosatlariga asoslangan DLP tizimidir. Ushbu yondashuvda xavfsizlik siyosatlari va qoidalar tashkilotning maxfiy ma’lumotlarini qanday boshqarish, kimga ruxsat berish va ma’lumotlarni qaysi sharoitlarda uzatish mumkinligini belgilaydi. - Content Inspection: Tarmoq orqali o‘tayotgan ma’lumotlarni tahlil qilish va maxfiy ma’lumotlarni yo‘qotishning oldini olish uchun ishlatiladigan algoritmlar. Content Inspection 36

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 (Mazmunni tekshirish algoritmlari) - axborotni yo‘qotishning oldini olish tizimlarining muhim qismidir. Ushbu algoritmlar ma’lumot oqimlari, fayllar va kommunikatsiyalarni tahlil qilib, noxosil foydalanish yoki siyosat buzilishlarini aniqlashga mo‘ljallangan[2]. Mazmunni tekshirish - bu faqat metadata yoki tashqi xususiyatlarga emas, balki ma’lumotlarning o‘ziga tahlil o‘tkazishdir. Ushbu algoritmlar hujjatlar, elektron pochta xabarlari yoki ma’lumot paketlaridagi haqiqiy mazmunni chuqur o‘rganib, oldindan belgilangan siyosatlarga zid axborotni aniqlaydi. Mazmunni tekshirish algoritmlari xususiyatlariga ko‘ra bir nechta turlarga bo‘linadi, ular orasida eng keng tarqalgan texnologiyalar sifatida quyidagilarni ko‘rsatish mumkin: - paketlarni chuqur tekshirish (Deep Packet Inspection, DPI); andoza aniqlash (Pattern Recognition); evriluvchan tahlil (Heuristic Analysis); tabiiy tilni qayta ishlash (Natural Language Processing, NLP); Fayllar izini aniqlash (File Fingerprinting). Deep Packet Inspection (DPI) internet orqali yuborilayotgan paketlarni chuqur tahlil qilish uchun qo‘llaniladigan texnologiyadir. Internet orqali amalga oshirilayotgan barcha aloqa jarayonlari ma’lumotlarni uzatish uchun "paketlar"dan foydalanadi. Bu jarayon VoIP qo‘ng‘iroqlarimiz, tashrif buyurayotgan veb-sahifalarimiz va yuborayotgan elektron pochta xabarlarimizni o‘z ichiga oladi. Pattern Recognition (shablonlarni tanish) - bu kompyuter ilm-fani va sun’iy intellekt sohasida keng qo‘llaniladigan texnologiya bo‘lib, u ma’lumotlar to‘plamlaridan shablonlar yoki qoidalarni aniqlashni nazarda tutadi. Shablonlarni tanish tizimlari ma’lum bir xususiyatlarga yoki parametrlar to‘plamiga asoslangan holda ma’lumotlarni tahlil qiladi va ularning ichidan o‘xshashliklarni aniqlaydi. Bu tizimlar ko‘pincha tasvirlar, ovozlar, matnlar va boshqa murakkab ma’lumotlar shakllarini qayta ishlashda ishlatiladi. Feedback / adaptation The real world Sensor Preprocessing and enhancement Feature extraction Classification algorithm Class assignment Clustering algorithm Cluster assignment Regression algorithm Predicted values 1-rasm. Shablonni tekshirish tizimlarining umumiy arxitekturasi Heuristic tahlil kiberxavfsizlik va antivirus protokollarining muhim tarkibiy qismi bo‘lib, asosan zararli dasturlar kabi jiddiy xavf-xatarlarga qarshi kurashish va ularni aniqlash maqsadida ishlatiladi. Bu usul yangi viruslar yoki hozirgi kunda tanilgan viruslarning varianti turlarini aniqlashda an’anaviy usullardan chetga chiqadi. Bu xavfsizlik dasturlariga qurilmaga zarar yetkazishi mumkin bo‘lgan noma’lum yoki xavfli faoliyatlarni aniqlash uchun zarur bo‘lgan ilgari ko‘rish imkonini beradi. 37

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot Tabiiy tilni qayta ishlash (NLP) - kompyuterlarni insonlarning tabiiy tili bilan ishlashga o‘rgatish uchun yaratilgan texnologiyalar va metodlar to‘plamidir. Bu soha, kompyuterlar va insonlar o‘rtasida tilni avtomatik tarzda tushunish va ishlashni ta’minlashga yo‘naltirilgan[3]. Fayl Fingerprinting (Fayl izini olish) - bu faylni unikal tarzda identifikatsiya qilish usuli bo‘lib, ma’lumotlar bazasidagi yoki tizimdagi fayllar orasida farqlash va ularga alohida e’tibor qaratish imkoniyatini beradi. Bu jarayon, asosan, faylni boshqalardan ajratib turadigan raqamli izni yaratishga asoslanadi. Faylning izini olish jarayonida uning tarkibi, xususiyatlari yoki boshqa identifikatsiyalovchi belgilariga asoslanib unikal raqamli identifikator hosil qilinadi. Axborot yo‘qolishi, ma’lumotlarning sizib chiqishi, ma’lumot o‘g‘irlanishi yoki hatto ma’lumot - dunyoning qaysi burchagida bo‘lishingizdan qat’i nazar tashkilotdan axborotning yo‘qolishini ifodalovchi atamalar mavjud. Dunyo bo‘ylab katta va kichik tashkilotlar o‘z ma’lumotlarini himoya qilish zarur. Buning sababi faqatgina axborot yo‘qolishi emas, balki bu holat qonunchilik, milliy jarimalar va hatto obro‘ga zarar yetkazishi mumkinligi bilan ham bog‘liq. DLP yechimlari ko‘p yillar davomida qo‘llanilib kelinmoqda, hatto ularning qo‘llanilishi endi kichik tashkilotlarga ham yetib borgan. Yangi texnologiyalarni an’anaviy DLP metodologiyalari bilan birlashtirish orqali hatto eng kichik biznes ham arzon narxda amalga oshirish oson bo‘lgan zamonaviy va murakkab yechimdan foydalanishi mumkin. Symantec Data Loss Prevention modullari va komponentlari taklif qilingan umumiy DLP arxitekturasiga osonlik bilan moslashtirilishi mumkin: 1. Data in Motion (Harakatdagi ma’lumotlar) Symantec DLP modullari: - DLP for Network: Bu modul tarmoq orqali ma’lumotlar oqimini nazorat qiladi va harakatdagi ma’lumotlarni aniqlash, bloklash yoki ogohlantirish kabi choralar ko‘radi. Tarmoq orqali o‘tadigan ma’lumotlarni, jumladan, elektron pochta, veb-pochta va boshqa Internet protokollarini tekshiradi[4]. 2. Data at Rest (Saqlangan ma’lumotlar) Symantec DLP modullari: - DLP for Storage: Bu komponent endpointlar, fayl serverlari, veb-portallar va ma’lumotlar bazalari kabi saqlash qurilmalaridagi ma’lumotlarni tekshiradi. 3. Data in Use (Foydalanishdagi ma’lumotlar) Symantec DLP modullari: 4. DLP for Endpoint: Bu modul foydalanuvchi qurilmalarida ma’lumotlar oqimini nazorat qiladi. Ushbu modul ma’lumotlarni bloklash, shifrlash, ogohlantirish, karantinlash yoki boshqa xavfsizlik choralarini ko‘radi. 38

Sun’iy Intellekt va Raqamli Iqtisodiyot SIRI-2025 Span Port MTA MTA Active Directory DLP DLP Network Network Monitor Monitor Oracle Oracle Database Database DLP DLP Enforce Enforce Server Server DLP DLP Network Network Discover Discover and and Prevent Prevent Administrator Administrator DLP DLP Endpoint Endpoint Discover Discover and and Prevent Prevent DLP DLP Network Network Prevent Prevent Email Email and and Web Web Proxy Proxy Data in Motion File Database File Share point point Database Share Server Server Server Server Server Server Data at Rest Web Email FTP Clipboard Print Data in Use 2-rasm. Semantik ma’lumotlar yo‘qotilishining oldini olish tizimining tavsiya etilayotgan arxitekturasi 5. Ma’lumotlarni Boshqarish va Tahlil qilish Symantec DLP modullari: - DLP Management Server: Bu markazlashtirilgan boshqaruv serveri siyosatlar yaratish va ularni boshqa DLP serverlariga yuborish, voqealarni va ish jarayonlarini boshqarish, xabarlarni yig‘ish va hisobotlarni taqdim etish uchun ishlatiladi. U boshqa uchinchi tomon tizimlari, masalan, Active Directory yoki SIEM bilan integratsiya qilinishi mumkin. 6. Ma’lumotlarning tarmoq orqali nazorati Symantec DLP modullari: - DLP for Web: Ushbu modul veb-aloqalarni nazorat qilish va ma’lumotlarni tarmoq orqali chiqishini aniqlash uchun ishlatiladi. Bu tizim ICAP protokoli yordamida DLP serveri bilan integratsiyalashadi va web-traffikni tahlil qiladi. Symantec DLP yechimi, arxitekturaning har bir qismiga mos ravishda, tashkilotlarga ma’lumotlarning holatini kuzatib borish va xavfsizlik choralarini joriy qilish imkoniyatini beradi. Boshqaruv konsoli orqali barcha modullarni bir joyda boshqarish va tahlil qilish mumkin[5]. Semantik Data Loss Prevention modullari va komponentlari taklif qilingan umumiy Data Loss Prevention arxitekturasiga osonlik bilan moslashtirilishi mumkin: Ushbu arxitektura yordamida ma’lumotlarning foydalanish holatini Symantec DLP Endpoint Discover va Protect komponentalarini joylashtirish orqali ta’minlaymiz. Saqlangan ma’lumotlar Symantec DLP Network Discover va Protect komponentalari orqali ta’minlanadi. Harakatdagi ma’lumotlar uchun esa Symantec DLP Network Monitor port-span yoki tarmoq tap orqali tarmoq bilan integratsiya qilinadi. 39

SIRI-2025 Sun’iy Intellekt va Raqamli Iqtisodiyot Xulosa Telekommunikatsiya tizimlarida ma’lumotlar xavfsizligini ta’minlash zamonaviy kiberxavfsizlik tahdidlariga qarshi samarali choralar ko‘rishni talab etadi. Tadqiqot natijalariga ko‘ra, Data Loss Prevention (DLP) algoritmlaridan foydalanish maxfiy ma’lumotlarning ruxsatsiz tarqalishi va o‘g‘irlanishining oldini olishda muhim ahamiyatga ega. Ayniqsa, Content Inspection va Policy-based DLP kabi yondashuvlar orqali tarmoq va qurilmalardagi ma’lumotlarni samarali himoya qilish imkoniyati yaratiladi. Symantec DLP yechimlari asosidagi taklif etilgan arxitektura esa real vaqtda nazorat, tahlil va boshqaruv imkoniyatlarini ta’minlab, tashkilotlarda ma’lumotlar xavfsizligini yanada oshiradi. Adabiyotlar 1. O‘zbekiston Respublikasi Prezidentining 05.10.2020 yildagi “Raqamli O‘zbekiston - 2030 strategiyasini tasdiqlash va uni samarali amalga oshirish chora-tadbirlari to‘g‘risida”gi PF-6079-son farmoni. 2. F.Pagani, M.Dell’Amico, D.Balzarotti Beyond precision and recall: understanding uses (and misuses) of similarity hashes in binary analysis // Proceedings of the Eighth ACM Conference on Data and Application Security and Privacy. – 2018. – С. 354-365 3. J.Oliver, C.Cheng, Y.Chen TLSH--a locality sensitive hash //2013 Fourth Cybercrime and Trustworthy Computing Workshop. – IEEE, 2013. – С. 7-13. 4. Vitor Hugo Galhardo Moia and Marco Aurélio Amaral Henriques Similarity Digest Search: A Survey and Comparative Analysis of Strategies to Perform Known File Filtering Using Approximate Matching // Security and Communication Networks Volume 2017, Article ID 1306802, 17 pages. 5. V. Gupta, "File detection in network traffic using approximate matching," MS Thesis. Institutt for Telematikk, 2013 40