黑料不打烊

艩iame magistro baigiamajame darbe tiriamas automatizuotas duomen懦 nutek臈jimo klasifikavimas, taikant turinio analize gr寞stus ma拧ininio mokymosi metodus. Nagrin臈jama problema did臈jantis poreikis identifikuoti ir skirstyti skirting懦 tip懦 nutek臈jusius duomen懦 寞ra拧us sud臈tingose technologin臈se aplinkose, kuriose tradiciniai taisykl臈mis pagr寞sti metodai tampa nepakankamais. Darbe si奴loma strukt奴rizuota klasifikavimo sistema, skirta duomen懦 pa啪eidimus skirstyti 寞 keturias kategorijas finansinius, asmens tapatyb臈s, medicininius ir 寞moni懦 duomen懦 nutek臈jimus remiantis pa膷i懦 寞ra拧懦 turiniu. Sukurta po啪ymi懦 i拧traukimo metodika, pagr寞sta PCI-DSS, HIPAA ir OWASP standartais, apimanti kredencial懦, tapatyb臈s, finansinius, medicininius ir 寞moni懦 duomenis atspindin膷ius po啪ymius. Keturi klasifikavimo metodai, i拧d臈styti did臈jan膷io sud臈tingumo tvarka, 寞gyvendinti ir sistemingai palyginti vienodomis eksperimentin臈mis s膮lygomis, vertinant tiek tikslum膮 su 拧variais duomenimis, tiek atsparum膮 duomen懦 degradacijos s膮lygomis. Gauti rezultatai rodo, kad med啪iais gr寞sti ansambli懦 metodai reik拧mingai lenkia linijinius ir taisykl臈mis pagr寞stus metodus, i拧laikydami auk拧t膮 klasifikavimo tikslum膮 net ir pablog臈jus duomen懦 kokybei. Daroma i拧vada, kad turinio pagrindu pagr寞stas duomen懦 nutek臈jimo klasifikavimas yra pakankamai patikimas praktiniam taikymui kibernetinio saugumo srityje.

This Master's Degree Thesis investigates the automated classification of data leaks using content-based machine learning methods. The central problem addressed is the growing difficulty of identifying and categorizing different types of leaked records in complex technological environments, where traditional rule-based approaches prove insufficient. The study proposes a structured classification framework that organizes data breaches into four categories financial, personally identifiable information, medical, and corporate based on the content of the leaked records themselves. A feature extraction pipeline is developed drawing on established security standards including PCI-DSS, HIPAA, and OWASP, capturing credential, identity, financial, medical, and corporate characteristics of each record. Four classification methods of progressively increasing complexity are implemented and systematically compared under identical experimental conditions, covering both clean data performance and robustness under realistic data degradation scenarios. The results demonstrate that tree-based ensemble methods significantly outperform linear and rule-based approaches, maintaining strong classification accuracy even when input data quality deteriorates substantially. The thesis concludes that content-based classification of data leaks is both technically feasible and reliable enough for practical deployment in cybersecurity operations.

Darbe tiriamas atvirojo kodo did啪i懦j懦 kalbos modeli懦 efektyvumas autonominiuose skvarbos testuose, pasitelkiant atvirojo kodo 寞rankius bei karkasus. Naudojant 鈥濸entestGPT鈥� 寞rank寞, realizuot膮 dirbtinio intelekto agento sprendimu, atliekami 鈥瀀BOW鈥� kibernetini懦 pa啪eid啪iamum懦 i拧naudojimo testai su skirtingais atvirojo kodo dirbtinio intelekto modeliais. Atlikt懦 test懦 rezultatai, kuriems i拧spr臋sti panaudoti atvirojo kodo dirbtinio intelekto modeliai, lyginami su u啪darojo kodo komercini懦 modeli懦 pasiekimais, analizuojant pa啪eid啪iamum懦 aptikimo efektyvum膮, i拧naudojimo realizacij膮 bei test懦 sprendinius skirtinguose sud臈tingumo lygmenyse bei pa啪eid啪iamumo kategorijose. Darb膮 sudaro 5 dalys: 寞vadas, dirbtinio intelekto modeli懦 ir agentini懦 sistem懦 analiz臈, metodologija ir 寞ranki懦 pasirinkimas bei sistemos architekt奴ros detalizacija, test懦, atlikt懦 su 3 skirtingais dirbtinio intelekto modeliais, rezultat懦 apibendrinimas ir palyginimas, i拧vados, literat奴ros s膮ra拧as. Darbo apimtis 鈥� 46 psl. teksto be pried懦, 36 iliustracijos, 8 lentel臈s, 31 literat奴ros 拧altinis.

The thesis investigates the effectiveness of open-source large language models in autonomous penetration testing, using open-source tools and frameworks. The PentestGPT tool, implemented as an AI agent solution, is used to conduct XBOW cybersecurity vulnerability exploitation tests with different open-source AI models. The results obtained using open-source AI models are compared against the achievements of closed-source commercial models through an analysis of vulnerability detection effectiveness, exploitation implementation and task outcomes across different difficulty levels and vulnerability categories. The thesis consists of 5 parts: introduction, analysis of AI models and agentic systems, methodology and tool selection with a detailed system architecture description, summary and comparison of test results obtained with 3 different AI models, conclusions, references. The thesis consists of 46 pages of text without appendices, 36 illustrations, 8 tables, and 31 references.

艩iame magistro darbe nagrin臈jamas WEB servis懦 API saugumas ir jo testavimo metodai. Darbe analizuojamos da啪niausiai pasitaikan膷ios API saugumo problemos, susijusios su autentifikavimu, prieigos kontrole, netinkama konfig奴racija, pa啪eid啪iamomis priklausomyb臈mis, 寞vesties duomen懦 apdorojimu ir i拧tekli懦 ribojimu. Atlikus literat奴ros analiz臋, pasi奴lyta strukt奴rizuota API saugumo testavimo metodika, apimanti funkcin寞 testavim膮, statin臋 programinio kodo saugumo analiz臋, dinamin臋 programinio kodo saugumo analiz臋, priklausomybi懦 ir komponent懦 skenavim膮, API fuzz testavim膮 bei apkrovos ir grei膷io ribojimo testavim膮. Eksperimentin臈je dalyje metodika pritaikyta pa啪eid啪iamoje API aplinkoje, siekiant 寞vertinti jos geb臈jim膮 aptikti skirting懦 tip懦 saugumo tr奴kumus. Tyrimo rezultatai parod臈, kad keli懦 testavimo metod懦 derinimas leid啪ia pla膷iau 寞vertinti API saugumo b奴kl臋. Nei vieno 寞rankio ar vieno testavimo tipo taikymas. Pasi奴lyta metodologija leid啪ia sistemingai rinkti, normalizuoti ir vertinti rezultatus pagal i拧 anksto apibr臈啪tus saugos rodiklius.

This master鈥檚 thesis investigates the security of web service APIs and methods used for their assessment. The work analyses common API security issues related to authentication, access control, insecure configuration, vulnerable dependencies, input handling and resource limitation. Based on the literature analysis, a structured API security testing methodology is proposed. The methodology combines functional testing, Static Application Security Testing, Dynamic Application Security Testing, dependency and component scanning, API fuzz testing, and load and rate-limiting testing. In the experimental part, the methodology is applied in a vulnerable API environment to evaluate its ability to detect different types of security weaknesses. The results show that combining several testing methods provides broader API security coverage than relying on a single tool or testing approach. The proposed methodology enables systematic collection, normalization and evaluation of results according to predefined security metrics. Therefore, it can be used as a repeatable approach for assessing API security from source-code, runtime, dependency, input-handling and resilience perspectives.

Bakalauro baigiamojo darbo tikslas 鈥� palengvinti darbuotoj懦 u啪imtumo ir darbo laiko apskaitos procesus, sukuriant darbuotoj懦 u啪imtumo informacin臋 sistem膮. Darbe atlikta darbuotoj懦 u啪imtumo ir darbo laiko apskaitos srities analiz臈, i拧nagrin臈tos pana拧ios paskirties informacin臈s sistemos bei j懦 k奴rimui taikomos technologijos. Remiantis atlikta analize suformuluoti funkciniai ir nefunkciniai reikalavimai, parengtas sistemos projektas, apimantis sistemos architekt奴r膮, duomen懦 model寞, klasi懦 diagram膮 ir naudotoj懦 s膮veikos modelius. Sukurta internetin臈 darbuotoj懦 u啪imtumo informacin臈 sistema leid啪ia valdyti darbuotoj懦 duomenis, registruoti ir tvirtinti neatvykim懦 pra拧ymus, generuoti bei tvirtinti darbo laiko tabelius, per啪i奴r臈ti darbo laiko apskaitos informacij膮 ir eksportuoti duomenis ataskaitoms. Sistema realizuota naudojant Python programavimo kalb膮, Django karkas膮 ir PostgreSQL duomen懦 baz臋. Sistemos kokybei 寞vertinti atlikti vienet懦, integraciniai ir E2E testai. Testavimo rezultatai parod臈, kad sistema atitinka suformuluotus reikalavimus ir gali b奴ti naudojama darbuotoj懦 u啪imtumo bei darbo laiko apskaitos proces懦 valdymui organizacijoje. Baigiam膮j寞 darb膮 sudaro 寞vadas, analitin臈 dalis, projektavimo dalis, realizavimo ir testavimo dalis, i拧vados, pasi奴lymai bei literat奴ros s膮ra拧as. Darbo apimtis 鈥� 59 puslapiai teksto, jame pateikti 15 paveiksl懦, 16 lenteli懦 ir panaudoti 19 bibliografini懦 拧altini懦.

The aim of this bachelor thesis is to simplify employee occupancy and working time accounting processes by developing an employee employment information system. The thesis includes an analysis of employee occupancy and working time accounting, a review of similar information systems, and an evaluation of technologies suitable for their development. Based on the analysis, functional and non-functional requirements were defined, and a system design was prepared, including system architecture, data model, class diagram, and user interaction models. A web-based employee employment information system was developed to manage employee data, submit and approve absence requests, generate and approve timesheets, review working time information, and export data for reporting purposes. The system was implemented using the Python programming language, the Django framework, and a PostgreSQL database. To evaluate the quality of the developed system, unit, integration, and end-to-end tests were performed. The testing results confirmed that the system meets the defined requirements and can be used for managing employee occupancy and working time accounting processes within an organization. The thesis consists of an introduction, an analytical part, a design part, an implementation and testing part, conclusions, recommendations, and a list of references. The thesis comprises 59 pages of text, contains 15 figures and 16 tables, and is based on 19 bibliographic sources.

Baigiamojo darbo tikslas 鈥� palengvinti ir pagreitinti 寞moni懦 darb膮 su dideliu dokument懦 kiekiu, sukuriant automatin臋 dokument懦 klasifikavimo sistem膮, pritaikyt膮 lietuvi懦 kalbai ir veikian膷i膮 lokaliai. Analitin臈je darbo dalyje buvo atlikta dokument懦 klasifikavimo metod懦, teksto analiz臈s technologij懦 ir pana拧i懦 sistem懦 analiz臈. Nustatyta, kad esami sprendimai da啪nai orientuoti 寞 vidutines ir dideles organizacijas, reikalauja debes懦 kompiuterijos paslaug懦 arba nepakankamai palaiko lietuvi懦 kalb膮. Taip pat i拧analizuoti 拧iuolaikiniai daugiakalbiai kalbiniai modeliai ir j懦 pritaikomumas dokument懦 klasifikavimo u啪daviniams. Palyginus technologinius sprendimus, serverio daliai buvo pasirinkta Python programavimo kalba ir FastAPI karkasas, naudotojo s膮sajai 鈥� React, TypeScript ir Electron technologijos, o semantinei analizei 鈥� intfloat/multilingual-e5-large modelis. Projektin臈je darbo dalyje suformuluoti funkciniai ir nefunkciniai reikalavimai, suprojektuota sistemos architekt奴ra, duomen懦 baz臈, naudotojo s膮saja bei dokument懦 klasifikavimo posistem臈. Nubrai啪ytos naudojimo atvej懦, komponent懦, sek懦 ir duomen懦 baz臈s diagramos. Realizacijos dalyje 寞gyvendinta lokaliai veikianti dokument懦 klasifikavimo sistema. Sistema leid啪ia registruoti ir autentifikuoti naudotojus, 寞kelti dokumentus, kurti ir valdyti 啪ymas, atlikti dokument懦 klasifikavim膮 semantinio pana拧umo bei nulinio 拧奴vio metodais, per啪i奴r臈ti klasifikavimo istorij膮 ir valdyti klasifikavimo rezultatus. Testavimo metu buvo atlikti vienetiniai ir rankiniai testai bei dokument懦 klasifikavimo tikslumo tyrimas, naudojant daugiakalb寞 agentlans/multilingual-document-classification duomen懦 rinkin寞. Testavimo rezultatai parod臈, kad sistema atitinka visus 12 funkcini懦 ir 10 nefunkcini懦 reikalavim懦, o vidutinis klasifikavimo tikslumas angl懦 ir lietuvi懦 kalbomis siek臈 69 %. Darbo apimtis: 64 psl., 28 pav., 13 lent., 33 拧altiniai.

The aim of this thesis is to facilitate and accelerate companies' work with large volumes of documents by developing an automated document classification system tailored to the Lithuanian language and operating locally. In the analytical part, an analysis of document classification methods, text analysis technologies, and similar systems was conducted. It was found that existing solutions are often oriented towards medium and large organizations, require cloud computing services, or lack sufficient support for the Lithuanian language. Modern multilingual language models and their applicability to document classification tasks were also analyzed. Comparing the technological solutions, Python and the FastAPI framework were chosen for the server side; React, TypeScript, and Electron for the user interface; and the intfloat/multilingual-e5-large model for semantic analysis. In the project part, functional and non-functional requirements were defined, and the system architecture, database, user interface, and document classification subsystem were designed. Use case, component, sequence, and database diagrams were created. In the implementation part, a locally operating document classification system was developed. The system allows users to register and authenticate, upload documents, create and manage tags, classify documents using semantic similarity and zero-shot methods, view classification history, and manage classification results. During testing, unit and manual tests were performed, along with a document classification accuracy study using the multilingual agentlans/multilingual-document-classification dataset. The test results showed that the system meets all 12 functional and 10 non-functional requirements, with an average classification accuracy of 69% both in English and Lithuanian. Thesis scope: 64 pages, 28 figues, 13 tables, 33 拧altiniai.

艩iame magistro baigiamajame darbe nagrin臈jamas Web API pa啪eid啪iamum懦 vertinimas ir ma啪inimas naudojant dirbtiniu intelektu paremt膮 sistem膮. Tyrime analizuojami da啪niausiai pasitaikantys 啪iniatinklio paslaug懦 API pa啪eid啪iamumai, taip pat j懦 aptikimui ir ma啪inimui naudojami 寞rankiai bei metodai. Darbo tikslas yra pagerinti Web API saugum膮, pasi奴lant metod膮, skirt膮 aptikt懦 pa啪eid啪iamum懦 vertinimui ir ma啪inimo rekomendacij懦 generavimui pasitelkiant dirbtin寞 intelekt膮. Sukurtas sprendimas apdoroja pa啪eid啪iamum懦 skeneri懦 ataskaitas, i拧skiria aktuali膮 informacij膮 apie pa啪eid啪iamumus ir naudoja OpenAI API kontekstin臈ms ma啪inimo rekomendacijoms generuoti. Vertinimas, atliktas naudojant realius testavimo atvejus, parod臈, kad si奴lomas dirbtiniu intelektu paremtas metodas gali pad臈ti programuotojams suprasti aptiktus pa啪eid啪iamumus ir pasirinkti tinkamus j懦 ma啪inimo veiksmus. Darb膮 sudaro 寞vadas, teorin臈 啪iniatinklio paslaug懦 API saugumo analiz臈, Web API pa啪eid啪iamum懦 skenavimas ir rezultatai, sukurto sprendimo 寞gyvendinimas, si奴lomos sistemos testavimas ir naudotojo s膮saja, bendrosios i拧vados ir literat奴ros s膮ra拧as. Darb膮 sudaro 73 puslapis teksto, 31 paveikslas, 11 lenteli懦 ir 21 bibliografinis 拧altinis.

This master thesis investigates the evaluation and mitigation of vulnerabilities in Web APIs using an AI-assisted system. The research analyzes common web service API vulnerabilities together with the tools and techniques used for their detection and mitigation. The thesis aims to improve Web API security by proposing a method for evaluating detected vulnerabilities and generating mitigation recommendations with the support of artificial intelligence. The developed solution processes vulnerability scanner reports, extracts relevant vulnerability information, and uses the OpenAI API to generate context-based mitigation recommendations. The evaluation based on real test cases shows that the proposed AI-assisted approach can support developers in understanding detected vulnerabilities and selecting appropriate mitigation actions. The thesis consists of the introduction, theoretical analysis of web service API security, web API vulnerability scanning and results, implementation of the developed solution, testing and user interface of the proposed system, general conclusions, and list of references. The thesis contains 73 pages of text, 31 figures, 11 tables, and 21 bibliographic sources.

艩is tyrimas skirtas duomenimis ir ontologija pagr寞st懦 metod懦 integravimui 寞 incident懦 valdymo proces膮. Ontologijos derinys su pa啪angia duomen懦 analitika padeda spr臋sti semantinio suderinamumo bei sprendim懦 pri臈mimo kokyb臈s problemas. Darbo tikslas - sukurti ir i拧bandyti nauj膮 model寞, leid啪iant寞 efektyviau valdyti incident懦 procesus bei didinti organizacijos lankstum膮. Atlikta sistemin臈 literat奴ros ap啪valga, identifikuoti BPM metodik懦 tr奴kumai ir pasi奴lytas sprendimas. Sukurtas modelis realizuotas Jira sistemoje naudojant ScriptRunner bei Prot茅g茅 aplinkoje apibr臈啪tas SWRL taisykles. Eksperimentiniai tyrimai ir simuliacijos patvirtino, kad ontologij懦 ir duomen懦 s膮veika statisti拧kai reik拧mingai padidina proceso efektyvum膮, suma啪ina L1 darbuotoj懦 u啪imtum膮 bei klaid懦 tikimyb臋. Darbas prisideda prie incident懦 valdymo tobulinimo, si奴lydamas inovatyv懦 po啪i奴r寞 寞 resurs懦 valdym膮 kintan膷iomis rinkos s膮lygomis.

This research is aimed at integrating data- and ontology-based methods into the incident management process. The combination of ontology with advanced data analytics helps to solve problems of semantic compatibility and decision-making quality. The aim of the work is to create and test a new model that allows for more effective management of incident processes and increase organizational flexibility. A systematic literature review was conducted, shortcomings of BPM methodologies were identified and solution was proposed. The created model was implemented in the Jira system using ScriptRunner and SWRL rules defined in the Prot茅g茅 environment. Experimental studies and simulations confirmed that the interaction of ontologies and data statistically significantly increases process efficiency, reduces L1 employee employment and the probability of errors. The work contributes to the improvement of incident management by offering an innovative approach to resource management in changing market conditions.

Potvyniai yra itin destruktyvi stichin臈 nelaim臈, turinti didel寞 ekonomin寞 ir socialin寞 poveik寞. Nors 拧iuolaikiniai ma拧ininio mokymosi modeliai pasiekia auk拧t膮 tikslumo lyg寞, j懦 nepermatomumas riboja modeli懦 interpretuojamum膮 鈥� veiksn寞, kuris yra ypa膷 svarbus priimant sprendimus. 艩iame tyrime si奴loma potvyni懦 prognozavimo sistema, pagr寞sta adaptyvi膮ja nerai拧kios logikos sistema (ANFIS). ANFIS yra interpretuojamas modelis, kuris derina neuroninio tinklo mokymosi geb臈jimus su taisykl臈mis pagr寞stu m膮stymu. Naudojant vie拧ai prieinamus Lietuvos duomenis, tyrime nagrin臈jamas 鈥瀊altosios d臈啪臈s鈥� ANFIS modelio potencialas pasiekti prognozavimo geb臈jimus, kurie prilygsta 鈥瀓uodosios d臈啪臈s鈥� metod懦 geb臈jimams, ypating膮 d臈mes寞 skiriant Minijos upei. Atliekama empirin臈 analiz臈, siekiant i拧tirti interpretuojamumo ir tikslumo kompromis膮. 艩iame tyrime taip pat daug d臈mesio skiriama modelio k奴rimui 鈥濸ython鈥� aplinkoje ir jo derinimui su kitomis technologijomis, pavyzd啪iui, LLM integracija rezultatams apibendrinti. Tyrime nagrin臈jamos sukurtos ANFIS modelio daugialyp臈s charakteristikos, atid啪iai i拧nagrin臈jant nerai拧kias taisykles ir 寞vestis. Toliau atliekama lyginamoji analiz臈, taikant klasikines ma拧ininio mokymosi metodikas, tokias kaip RNN, 鈥瀀GBoost鈥�, LR ir SVR, siekiant 寞vertinti si奴lomo modelio veiksmingum膮 ir na拧um膮. 艩io tyrimo rezultatai prisideda prie paai拧kinamojo ma拧ininio mokymosi srities pl臈tros, nes juose vertinamas skaidri懦 modeli懦 potencialas u啪tikrinti patikim膮 ir veiksming膮 potvyni懦 prognozavim膮. Be to, tyrimo rezultatai pritaikomi praktikoje: sukurtas modelis 寞diegiamas realaus laiko programoje, kuri kas valand膮 atnaujina potvyni懦 prognozes. Modelis taip pat pritaikytas naudoti kitoms did啪iosioms Lietuvos up臈ms. Strukt奴ra: 寞vadas, susijusi懦 darb懦 analiz臈, potvyni懦 prognazavimo sistemos metodo pasi奴lymas, potvyni懦 prognazavimo metodo k奴rimas. i拧vados, literat奴ra ir vienas priedas. Darb膮 sudaro 107 p. teksto be pried懦, 20 paveiksl懦, 17 lentel臈s, 55 bibliografini懦 寞ra拧懦.

Flooding remains a highly destructive natural disaster, with significant economic and social impacts. Whilst contemporary machine learning models attain a high level of accuracy, their opaque nature curtails the interpretability of the models, a factor which is critical in decision-making. The present study proposes a flood prediction system based on the Adaptive Neuro-Fuzzy Inference System (ANFIS). The ANFIS is an interpretable model that combines the learning capability of a neural network with rule-based reasoning. Using publicly accessible Lithuanian data, the research investigates the potential of a white-box ANFIS model to attain predictive capabilities that are commensurate with those of black-box approaches, with a particular focus on the MInija river. An empirical analysis is conducted for the purpose of examining the trade-off between interpretability and accuracy. The present study focuses on developing the model in the Python environment and combining it with other technologies, such as LLM integration for the results summary. The study investigates the multifaceted characteristics of the constructed ANFIS model through the examination of the fuzzy rules and inputs. A comparative analysis is then conducted with typical machine learning methodologies, including RNN, XGBoost, LR and SVR, to ascertain the efficacy of the proposed model. The findings of this study contribute to the field of explainable machine learning by assessing the potential of transparent models to provide reliable and effective flood forecasting. Furthermore, the findings of the study are applied in a practical context by deploying the developed model into a real-time application that updates flood forecasts on a hourly basis. The model has also been adapted for use with other major Lithuanian rivers. Structure: introduction, related work analysis, proposed approach on flood forecasting system, implementation of the proposed approach, conclusions, references and one appendix. Thesis consist of 107 p. text without appendixes, 20 figures, 17 tables, 55 bibliographical entries.

艩iame magistro baigiamajame darbe nagrin臈jamas kibernetini懦 incident懦 valdymo koordinavimo i拧拧奴kis saugumo operacij懦 centruose (SOC), kuriuose da啪nai susiduriama su persp臈jim懦 pervargimu, 寞ranki懦 fragmentacija ir rankini懦 darbo proces懦 neefektyvumu. Remiantis projektavimo mokslo tyrimo metodologija, atlikta lyginamoji atvirojo kodo incident懦 valdymo platform懦 鈥濼heHive", 鈥濿azuh" ir FIR analiz臈, kurios pagrindu pasirinkta ir adaptuota greitojo incident懦 reagavimo (FIR) platforma. Suprojektuoti ir 寞gyvendinti penki automatizavimo procesai: automatizuotas persp臈jim懦 i拧ankstinis apdorojimas su dideli懦 kalbos modeli懦 (LLM) pagrindu veikian膷iu klaiding懦 teigiam懦 rezultat懦 slopinimu, dinaminis pamainos principu veikiantis incident懦 priskyrimas, DI generuojamos pamainos perdavimo ataskaitos, kriptografinis 寞rodym懦 mai拧ymas (SHA-256) ir vieninga incident懦 laiko juosta su realiojo laiko bendradarbiavimu. Sukurtas prototipas 寞diegtas naudojant 鈥濪ocker Compose" ir 寞vertintas simuliuot懦 SOC scenarij懦 pagrindu. Rezultatai parod臈, kad automatizuotas priskyrimas suma啪ino reagavimo laik膮 90 %, klaiding懦 persp臈jim懦 slopinimas suma啪ino analitik懦 apkrov膮 85 %, o pamainos perdavimo trukm臈 sutrump臈jo nuo 15 minu膷i懦 iki ma啪iau nei 2 minu膷i懦. Sistema u啪tikrina teismin寞 vientisum膮 per kriptografin寞 寞rodym懦 grandin臈s saugojim膮 ir duomen懦 privatum膮 per vietin寞 LLM i拧vad懦 atlikim膮, sudarydama ekonomi拧kai efektyvi膮 alternatyv膮 SM漠, negalin膷ioms 寞sigyti komercini懦 SOAR sprendim懦

This study focuses on addressing the problem of cyber incident coordination in SOCs by mitigating issues such as alert fatigue, tool fragmentation, and inefficient workflows. By employing the Design Science Research Methodology, a comparative review of existing free-to-use incident management systems, TheHive, Wazuh, and FIR led to the choice of the Fast Incident Response (FIR) system. As a result, five processes were developed and integrated into the chosen system, including automated pre-processing of alerts based on the use of the language learning model to suppress false positives, dynamic distribution of incidents according to shifts, shift summarization based on artificial intelligence technology, cryptographic hash generation using SHA-256, and integration of a unified timeline along with real-time communication capability within the system. Furthermore, an adapted FIR system was containerized using Docker Compose and tested in a series of simulated SOCs scenarios. As a result, there was an 90% reduction in the time taken to assign analysts due to automation, an 85% reduction in the analyst's alert load as a consequence of the suppression of false-positive alerts, and shift summarization time was reduced from 15 minutes to less than 2 minutes.

艩iame magistro baigiamajame darbe si奴loma tobulinti programin臈s 寞rangos reikalavim懦 specifikacijos (SRS) proces膮 bankini懦 pokalbi懦 robot懦 sistem懦 kontekste, taikant dirbtiniu intelektu pagr寞stus SRS metodus. Kaip 啪inoma, tradiciniai metodai susiduria su sunkumais prisitaikant prie dinami拧kos, duomenimis grind啪iamos ir nedeterministin臈s dirbtinio intelekto sistem懦 prigimties, kai naudotoj懦 poreikiai, sistemos elgsena ir 啪ini懦 拧altiniai laikui b臈gant gali nuolat keistis. 艩iame darbe si奴lomas paie拧ka papildytos generacijos (angl. *Retrieval-Augmented Generation*, RAG) metodas, kuris sujungia semantin臈s paie拧kos, nat奴raliosios kalbos apdorojimo ir bankininkyst臈s srities specifin臈s 啪ini懦 baz臈s (KB) ypatybes su did啪iuoju kalbos modeliu pagr寞stu funkcini懦 reikalavim懦 (FR) generavimu. Sukurtame prototipe RAG 寞gyvendinamas naudojant Sentence-BERT u啪klaus懦 寞terpiniams kurti, FAISS semantinio pana拧umo paie拧kai ir kalbos model寞 strukt奴ruotiems funkciniams reikalavimams generuoti standartine forma 鈥濻istema turi...鈥�. Metodas taip pat 寞gyvendintas kartu su 啪ini懦 sprag懦 aptikimo ir gr寞啪tamojo ry拧io funkcijomis, kurios leid啪ia sistemos mokytojui identifikuoti tr奴kstam膮 arba nei拧sami膮 informacij膮 apie pokalbi懦 roboto 啪inias ir imtis atitinkam懦 veiksm懦. 艩is metodas buvo i拧bandytas naudojant bankinio pokalbi懦 roboto duomen懦 rinkin寞. Rezultatai rodo, kad 拧is metodas gali palengvinti ai拧kiai strukt奴ruot懦 ir kontekstui aktuali懦 funkcini懦 reikalavim懦 generavim膮, taip pagerindamas SRS proces膮. Ta膷iau tolesni tyrimai reikalingi siekiant geriau atskirti funkcinius reikalavimus nuo nefunkcini懦 reikalavim懦.

In this master's thesis, the improvement in Software Requirement Specification (SRS) process with respect to banking chatbot systems is suggested by implementing AI-enabled SRS techniques. As is well known, traditional techniques face issues in accommodating dynamic, data-centric and non-deterministic nature of AI systems wherein needs of users, behaviour of the system, and sources of knowledge may keep on changing over time. In this thesis, Retrieval-Augmented Generation (RAG) based approach combining features like semantic retrieval, natural language processing and banking domain specific Knowledge Base (KB) augmented with large language model based functional requirements (FR) generation has been proposed. In the developed prototype, the implementation of RAG is done using Sentence-BERT for query embedding, FAISS for semantic similarity retrieval and language model to generate structured FR in standard "The system shall..." form. The approach has been implemented along with knowledge gap detection and feedback facility which enables system trainer to identify missing/incomplete information about the chatbot knowledge and take appropriate actions. This approach has been tested using banking chatbot dataset. Results show that the approach is capable of facilitating generation of clearly structured and contextually relevant FRs, thus improving the process of SRS. However, further research is required in distinguishing between FRs and non-functional requirements.