Sirvi Märksõna "Artificial Intelligence" järgi

Nüüd näidatakse 1 - 17 17

listelement.badge.access-status Avatud juurdepääs ,
A Survey of Machine Learning Methods and their Applicability for Security Analysis
(Tartu Ülikool, 2024) Šramova, Diana; Matulevičius, Raimundas, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Selles lõputöös esile tõstetud probleem on määrata, milliseid Masinõppe ja Süvaõppe meetodeid tuleks infotehnoloogia (IT) turvaohtude tuvastamiseks rakendada. Kuna IT-turberünnakuid on praeguse tehnoloogia ja ressurssidega üha raskem tuvastada, vajavad tänapäevased tuvastussüsteemid lahendusi, mis kasutavad tehisintellekti alamhulki töökindluse ja automatiseerimise tagamiseks. Lahendus selle probleemi lahendamiseks on Masinõppe ja Süvaõppe meetodite analüüs, mis hindab nende rakendatavust kolmel turbejuhtumil: kasutajate ja üksuste käitumise analüütika tuvastamine, haavatavuse tuvastamine, ja andmepüügi tuvastamine. See analüüs hõlmab nii järelevalvega kui ka järelevalveta meetodeid, sealhulgas otsustusmetsa, tugivektormasinaid, logistilist regressiooni, K-lähimaid naabreid, klasterdamist, assotsiatsioonireegleid, korduvaid neurovõrke, konvolutsioonilisi neurovõrke, virnastatud autokodeerijaid ja vastandlikke generatiivseid võrke. Neid meetodeid käsitletakse konkreetsete turvajuhtumite sisendite, väljundite, tugevate ja nõrkade külgede põhjal. See õppemeetod tagab klassifitseerimise, mustrite äratundmise, anomaaliate tuvastamise ja läbitungimise testimise, suurendades turvasüsteemide töökindlust ja automatiseerimist. See lahendus annab turbespetsialistidele juhiseid Masinõppe ja Süvaõppe tehnikate valimiseks, mida mida tuleks rakendada konkreetsete IT-turbeülesannete puhul, vähendades seeläbi riske ja leevendama turbeohtusid.
listelement.badge.access-status Avatud juurdepääs ,
Anomaly Detection in CDR-Based Trajectories of the Mobile Cellular Network
(Tartu Ülikool, 2024) Vaino, Tiit; Hadachi, Amnir, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Mobiilse kõne- ja andmeside andmestik on suurepärane viis inimeste asukoha ja liikumise ligikaudseks hindamiseks. Seda saaks eetiliselt kasutada nii valitsuste kui ka eraettevõtete poolt. Näiteks kasutavad hädaabiteenused inimeste telefonide asukohti, et teada saada, kuhu saata abi. Ligikaudne asukoht saadakse mobiili ja mobiilivõrgu andmete kombineerimisega mobiilivõrgu kärgede (mobiilimasti antennide katvusalade) asukohtadega. Mitmed asukoha sündmused koos järjestikuste ajatemplitega saab kokku panna üheks trajektooriks. Neid asukohti ja trajektoore võiks kasutada mitmesuguste valitsuse või ärivaldkonna küsimuste analüüsimiseks. On oluline tagada, et andmekogumid oleksid puhtad, et vältida vigaste analüüside põhjal kallite vigade tegemist. Põhiline probleem on selles, et kärje asukoht andmebaasis ja reaalses elus ei kattu, sest on tehtud inimlik viga või andmebaaside pole sünkroonis. Antud magistritöö pakub välja mudeli anomaaliate tuvastamiseks CDR-põhistes trajektoorides, kasutades Trajectory Anomaly Detection with Mixed Feature sequence (TAD-FM) ehk siis trajektoori anomaaliate tuvastamine kombineeritud tunnusjoonte jadaga lähenemisviisi. Mudeli treenimine ja testimine viidi läbi reaalsete andmetega, kuhu olid integreeritud virtuaalsed anomaaliad, kus mõne kärje asukohti oli tahtlikult muudetud. Lisaks on mudelile tehtud täiustusi, et vähendada treenimise ja ennustamise ajalist keerukust. Pakutav mudel suutis märgistada ja tuvastada 66% kärgedest, millel olid valed asukohaandmed, kui kõrvalekalded.
listelement.badge.access-status Avatud juurdepääs ,
Asset-Oriented Threat Analysis for Large Language Model Systems
(Tartu Ülikool, 2025) Karagjaur, Mihhail; Matulevičius, Raimundas, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Large language model (LLM) deployments continue to proliferate across enterprises without systematic guidance on risk analysis of the LLM-based systems. Addressing this gap, the present study designs and validates an asset-oriented threat model, tailored to LLM systems. The research follows a design-science research paradigm. The research method incorporates (1) a systematic literature review of 45 peer-reviewed and grey sources, which led to the definition of 13 parent attack classes, a total of 24 threat variants. (2) A design of a threat model, which formalized the LLM business and system assets, their security criteria, mapped threats, security requirements, and countermeasures. (3) Two validation procedures, comprising a feasibility analysis of the threat model’s applicability and an empirical test of a jailbreak attack. The feasibility analysis determined that the proposed threat model, mapped to the Mistral Small 3.1, achieved a completeness score of 0.93 out of 1.00. Thus, indicating all but one of the seven system assets were fully represented in the real-world system. To further substantiate the applicability of the threat model, a jailbreak attack (prompt-injection) was executed with 100 prompts from the JailbreakV-28K benchmark open dataset. Without an official safety measure enabled, 78% of applicable prompts resulted in harmful output. With the safety measure enabled, the rate of harmful output was reduced to 70%. Indicating partial but insufficient mitigation. The main artifact of the thesis is an asset-oriented threat model for LLMs. The artifact consists of the following components: 1. High-level UML class and state, and BPMN process diagrams, depicting an LLM system, and mapping elicited threats to the system’s assets. 2. An interactive web page, which allows practitioners to traverse the produced threat model and to acquire information about the elicited assets, threats, and proposed countermeasures. 3. Code of the interactive web page, empirical tests, and datasets, supporting local use of the threat model and reproducibility of the jailbreak empirical test. Findings conclude that the LLM system possesses a wide attack surface while adding unique vectors such as jailbreak and embedding inversion. The thesis provides security and AI engineers with a systematic approach to risk analysis and countermeasure selection. Although the threat model was validated on a single open-weight model, the baseline methodology is model-agnostic and extensible. Future studies could validate the threat model against a wide set of LLM systems and automate control recommendations in the scope of DevSecOps.
listelement.badge.access-status Avatud juurdepääs ,
Emergent Theory of Mind (ToM) from Token Merging
(Tartu Ülikool, 2023) Tik, Chan Wai; Roy, Kallol, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Suured keelemudelid, kasutades süvaõppe meetodeid, annavad üllatavaid tulemusi inimtasemel jõudluse saavutamisel mitmesugustes ülesannetes, eriti see näitab tehisintellekti (AGI) võimekust ja valeuskumuste mõistmist ainult keeleõppest. See viitas tugevale seosele vaimuteooria (ToM) ja keeleõppe vahel, mida on psühholoogia valdkonnas kaua uuritud. See tõstatab küsimuse, kas see on sama ka nende süvaõppe keelemudelite puhul. Selles suunas keskendub lõputöö loomuliku keele mõistmise ülesandele (semantiline klassifikatsioon) ja esitab hüpoteesi, et peale statistika põhineb keelemudeli mõistmine koolitusest tulenevale süvastruktuurile. See struktuur ei ole aga otseselt juurdepääsetav, vaid ainult testide kaudu, nagu ka psühholoogia ToM-i uuringud. Seega on käesolevas lõputöös välja pakutud meetod – Token Merge, mis võimaldab testida struktuuri olemasolu. Eksperimendi tulemus annab positiivset tagasisidet väljapakutud hüpoteesi toetamise kohta ning annab ka järjestuse soorituse tähtsuse kohta grammatilise märgistamise teel.
listelement.badge.access-status Avatud juurdepääs ,
Facial Expression Recognition using Neural Network for Dyadic Interaction
(Tartu Ülikool, 2020) Sham, Abdallah Hussein; Ozcinar, Cagri; Tikka, Pia; Anbarjafari, Gholamreza; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Tehnoloogiainstituut
Computers are machines that don’t share emotions as humans do. With the help of Machine Learning (ML) and Artificial Intelligence (AI), social robots can become a reality. These robots are currently capable of interacting with people at a certain level, but not exactly as a person would do. For them to reach that level, they would need to understand more about how people interact daily and to learn from the dyadic interaction of two people would be a good option. Participants’ facial expressions are the main features that can be retrieved from dyadic interaction and this can be done using a trained Deep Neural Network (DNN) model. The DNN model, known as the Mini-Xception, is trained in this thesis using a dataset that has been pre-processed and can then be tested on images. Using a face detector algorithm, the model will be able to detect a person’s facial expression on the image. After successful image results, the model can be tested using a different medium. First, the tests are carried out using a webcam, then videos with more than one participant. Since people react to expressions, their reactions can also be caused by a context in which, for example, sad news would be the reason for sad emotion. The results of the tests will, therefore, be used for analysis where a correlation can be constructed between facial expressions and context.
listelement.badge.access-status Avatud juurdepääs ,
Facilitating the Automation of Use Case Specifications and Test Case Generation by Developing an LLM-Powered Tool
(Tartu Ülikool, 2025) Meus, Karl-Andreas; Georgiadis, Marinos, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Software projects are often delayed or experience increased costs due to unclear or incomplete requirements. This thesis addresses these challenges by introducing ReqFlowly, a web-based application that automates key aspects of the requirements engineering process. ReqFlowly utilizes a Large Language Model (LLM), selected through a comparative evaluation. The model is guided by a formal use case grammar and structured prompts to help users transform initial stakeholder requirements into structured artifacts: domain entities, precise use case specifications, and comprehensive test cases. The proposed solution aims to minimize manual effort during the early phases of software development. It facilitates the production of software specifications that are more complete, consistent, and precise, thereby enhancing the overall efficiency and reliability of the engineering process
listelement.badge.access-status Avatud juurdepääs ,
Generative AI in Data Quality Management
(Tartu Ülikool, 2025) Roht, Karen; Nikiforova, Anastasija, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Kvaliteetsed andmed on teadlike ja täpsete otsuste langetamiseks hädavajalikud, kuid kvaliteedi saavutamine ja säilitamine võib olla keeruline ja ressursimahukas protsess. Generatiivsel tehisintellektil on potentsiaal andmekvaliteedi halduse hõlbustamises ja käsitsi töö minimeerimises, võimaldades protsesse automatiseerida. Lõputöö uurib, kuidas saaks kasutada generatiivset tehisintellekti andmekvaliteedi halduse protsessides nende väljakutsete lahendamiseks. Töös esitletakse tüpoloogiat, mis töötati välja tuginedes turul saadaolevate andmekvaliteedi tööriistade süstemaatilisele analüüsile ja mida seejärel valideeriti ning täiustati kaasates andmekvaliteedi eksperte. Ainult 13 tööriista analüüsitud 88-st kasutavad praegu generatiivset tehisintellekti. Kuigi eksperdid nägid generatiivse tehisintellekti potentsiaali rutiinsete ülesannete automatiseerimises ja skaleeritavuse parandamises, väljendati kahtlusi usaldusväärsuse, selgitatavuse ja andmete turvalisuse pärast. Seetõttu on oluline säilitada inimesepoolne järelvalve ja kasutada generatiivset tehisintellekti toetava tööriistana, mitte iseseisva otsustajana. Arvestades teadusuurimuste piiratud arvu ja generatiivse tehisintellekti madalat kaasatust andmekvaliteedi tööriistadesse, annab töö ülevaate generatiivse tehisintellekti kasutusvõimalustest ja rõhutab edasiste uurimuste ja vastutustundliku rakendamise vajadust.
listelement.badge.access-status Avatud juurdepääs ,
Heat Pump Detection from Household Electricity Consumption Using Different Machine Learning Classifiers
(Tartu Ülikool, 2022) Mukose, John-William; Roy, Kallol, juhendaja; Eljand, Kristjan, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Heat pumps are becoming more popular for heating houses as they use less energy than traditional heating methods, offsetting their higher installation costs. They have the added advantage of causing less pollution, due to less energy consumption and due to the ability to use electricity generated from renewable sources. This thesis aims to determine whether certain premises have heat pumps installed or not, based on their hourly electricity consumption. This is a time series classification task i.e. the hourly electricity consumption of each household in the dataset is a time series, and a classifier is to be trained on this data to be able to classify a household as having a heat pump installed or not. Different machine learning models are used: Recurrent and Convolutional Neural Networks, as well as Logistic Regression.The latter serves as a baseline to compare deep neural models against simple, yet interpretable logistic regression models. The ground truth data as to whether a premise has a heat pump or not is obtained from Eesti Energia heat pump sales records. We face an additional challenge of training the machine learning models with a small dataset of only 113 premises with heat pumps. We found that 2D Convolutional Neural Network with time series data reshaped into a 2 Dimensional image is the optimum classifier for our data. Our thesis presents an innovative solution of using CNN on heat pump time-series data instead of using sequential models of Long Short Term Memory (LSTM) networks, which are normally the main model used for time series data. In this case, the CNN has the advantage over LSTM of faster training times as well as better accuracy. The results come with the caveat that better and more reliable results can be obtained if a larger dataset becomes available.
listelement.badge.access-status Avatud juurdepääs ,
How AI is used in Digital Marketing
(Tartu Ülikool, 2021) Golubovic, Danijel; Milani, Fredrik Payman, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
There is a recognized need for Artificial Intelligence and Digital Marketing, and their popularity has been noticed in the past decade. Companies have raised significant interest in the development of AI implementation in Digital Marketing. Furthermore, they strive to explore which Artificial Intelligence algorithms are the most effective within Marketing strategy. In addition to this, big data become a central figure of investment toward data-driven decisions in companies. This is due to empirical findings of how the company's users behave, their actions, and how to boost revenue by processing such data. If we take into consideration how worldwide, multinational companies are putting tremendous resources into R&D departments to go ahead of the competition in terms of AI-powered Digital Marketing, it can be said that traditional advertising will be replaced in full capacity by 'intelligent' Digital Marketing campaigns. Thus, this study investigates which AI approaches are applicable in Digital Marketing, which Digital Marketing aspects are most flexible to be under AI impact and how AI can shape the Digital Marketing process to evolve as a better version of itself. The outcome of this thesis can be a good base for further research about combin-ing Digital Marketing and AI, and also it can contribute to other persons of interest in the profes-sional sense.
listelement.badge.access-status Avatud juurdepääs ,
Machine Learning Methods in Anti-Money Laundering
(Tartu Ülikool, 2025) Malkovski, Anton; Aktas, Kadir, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Anti-Money Laundering (AML) is a critical operation in the financial sector, and with the constant growth in transaction volume, traditional AML methods are no longer sufficient for effectively detecting and preventing money laundering activities. Machine learning (ML) has the potential to discover complex patterns within the vast amounts of transactional data and reduce the false positives (FP) in the AML alerts. This thesis analyzes the applicability of machine learning in AML and proposes a full training pipeline that covers model training, hyperparameter optimization, and synthetic data generation. The work focuses on training a machine learning model with focus on reducing FP noise while being able to classify true positive (TP) alerts by augmenting the highly imbalance training dataset with synthetically generated minority class samples using a Variational autoencoder (VAE) and applying Optuna hyperparameter optimization to tune the model. The results of the experiments demonstrate that this method improves the model’s performance while maintaining its ability to generalize to unseen data, finally achieving noise reduction of FP alerts by 40%.
listelement.badge.access-status Embargo ,
Public Encounters and Government Chatbots: When Servers Talk to Citizens
(Wiley, 2025-07-04) Alishani, Art; Homburg, Vincent; Velsberg, Ott
Public service providers around the world are now offering chatbots to answer citizens' questions and deliver digital services. Using these artificial intelligence-powered technologies, citizens can engage in conversations with governments through systems that mimic face-to-face interactions and adjust their use of natural language to citizens' communication styles. This paper examines emerging experiences with chatbots in government interactions, with a focus on exploring what public administration practitioners and scholars should expect from chatbots in public encounters. Furthermore, it seeks to identify what gaps exist in the general understanding of digital public encounters.
listelement.badge.access-status Avatud juurdepääs ,
Reaalajas objekti valimine, jälgimine ja asendamine taustaga
(Tartu Ülikool, 2025) Made, Rasmus; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
With the development of deep learning-based models, image and video segmentation has made a significant leap forward. The aim of this bachelor's thesis was to create a real-time video stream segmentation demo that detects a person in the image and replaces them with background pixels. During the work, various modern deep learning-based segmentation models were tested, and the most suitable ones were selected for the final demo. The purpose of the completed demo was to introduce the capabilities of modern segmentation models to Delta visitors. An analysis based on user feedback revealed that for the demo to function more effectively, a computer with high hardware capacity is required.
listelement.badge.access-status Avatud juurdepääs ,
Systematic Evaluation of Trustworthy AI Augmentation in Modern Applications
(Tartu Ülikool, 2024) Marasinghe, Marasinghe Mudiyanselage Rasinthe; Flores, Huber, juhendaja; Ottun, Abdul-Rasheed, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Tehisintellekt (AI) on levinud mitmetes sektorites, sealhulgas tervishoius, rahanduses, hariduses ja transpordis, muutes ülesannete täitmist ja otsuste tegemist. Kuid tehisintellekti kiire integreerimine on tekitanud märkimisväärset muret eraelu puutumatuse, eelarvamuste, turvalisuse ja tehisintellekti süsteemide, mida sageli nimetatakse „mustadeks kastideks“, läbipaistmatuse pärast. Need probleemid rõhutavad kriitilist puudujääki, mis on seotud tehisintellekti süsteemide tõhususe ja usaldusväärsuse tagamisega. Käesolevas uuringus käsitletakse seda puudujääki, keskendudes pideva inimliku järelevalve praktilisele rakendamisele tehisintellekti arendamisel. Uuringus hinnatakse konkreetselt SPATIAL platvormi jaoks välja töötatud adaptiivset armatuurlauda, et suurendada tehisintellekti läbipaistvust ja aruandekohustust. Uuringus hinnatakse eksperimentide kaudu meditsiinilise analüüsimooduliga (MAM), mis kasutab seletava tehisintellekti (XAI) meetodeid, et anda elektrokardiogrammi (EKG) andmeid analüüsivatele sidusrühmadele rollipõhiseid selgitusi, tehisintellekti genereeritud selgituste tõlgendatavust ja süsteemi toimimist erineva kasutajakoormuse korral. Tulemused näitavad, et kohandatud selgitused parandavad märkimisväärselt kasutajate arusaamist ja usaldust, samas kui süsteem säilitas tugeva jõudluse, tagades skaleeritavuse ja usaldusväärsuse. Need teadmised annavad väärtuslikke juhiseid praktiliste vahendite väljatöötamiseks, et tõhustada tehisintellekti järelduste jälgimist ja järelevalvet, mis on kooskõlas usaldusväärse tehisintellekti regulatiivsete nõuetega.
listelement.badge.access-status Avatud juurdepääs ,
Text-Driven Weakly Supervised Medical Image Segmentation
(Tartu Ülikool, 2025) Injarabian, David Avedis; Fishman, Dmytro, juhendaja; Ariva, Joonas, juhendaja; Šuvalov, Hendrik, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Medical image analysis has become an essential tool for clinical diagnosis, enabling specialists to detect, segment, and monitor various pathologies. Convolutional neural networks have traditionally dominated this field, achieving success in classification and segmentation tasks by relying solely on visual patterns. However, due to their inherent architectural limitations, they are unable to effectively incorporate complementary information such as textual reports provided by medical professionals. Recently, multimodal models, particularly Transformer-based vision-language architectures, have demonstrated promising results in general image recognition and generation tasks by effectively integrating text and visual data. Despite these advances, the potential of multimodal approaches in medical imaging, especially in the context of complex 3D volumetric data such as computer tomography scans, remains largely unexplored. This thesis investigates whether textual context provided by radiology reports can implicitly guide multimodal models to learn spatial locality in their counterpart medical images, potentially leading to emergent segmentation capabilities without explicit segmentation supervision. Such an approach could address the chronic shortage of manually annotated segmentation data, as obtaining these labels is expensive and labor-intensive. By examining how multimodal models trained on paired 3D computer tomography scans and radiology reports respond to textual prompts, the thesis seeks to understand if these models inherently learn meaningful spatial relationships. If multimodal models demonstrate implicit segmentation capabilities, they could serve as a valuable source for generating synthetic weakly supervised segmentation masks, reducing the need for costly manual annotation and supporting radiologists in clinical interpretation and triage workflows.
listelement.badge.access-status Avatud juurdepääs ,
Towards AI for cloud services reliability using combined metrics
(Tartu Ülikool, 2020) Chhetri, Tek Raj; Srirama, Satish Narayana, juhendaja; PhD Dehury, Chinmaya, juhendaja; MSc Lind, Artjom, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
With the emergence of cloud computing and the Quality of Services (QoS), Compute Power, Performance, and Scalability it offers, the paradigm of computing has shifted towards the cloud. Due to attractiveness cloud offers, today, more and more businesses, research, and individuals are adopting cloud services. Even with the maturity of the cloud, reliability is still a concern. The reason being the constant occurrence of failure causes financial loss as well as a negative impact on its users as it directly affects QoS. Further, the scale and heterogeneity make it more prone to failure, highlighting the necessity for a robust solution to maintain the attractiveness and prevent financial loss. By predicting failure before it could happen, we can improve the reliability. Artificial Intelligence, now, has made significant progress, finding itself a place in all possible areas. In our study we present artificial intelligence with a combined metrics approach to improve the failure prediction. An experiment conducted with data recorded from more than 100 cloud servers shows significant improvement in failure prediction with high prediction accuracy, precision, and recall compared to state of the art studies.
listelement.badge.access-status Avatud juurdepääs ,
When citizens meet the chatbot: Evidence from a survey vignette experiment in Estonia
(Sage Publications, 2025-12-10) Alishani, Art; Homburg, Vincent
Governments are increasingly using chatbots to facilitate public encounters. Using survey vignette experiments conducted in Estonia, we conclude that citizens' perceived usefulness and trust in technology significantly correlate with the intended use of chatbots in public encounters. Privacy concerns are related to the intended uses of chatbots for service provision but not for information provision, while trust in government, explainability, and the amount of information are not related to the intended uses of chatbots. These findings contribute to a better understanding of how citizens’ attitudes and perceived risks are relevant to citizens’ intended use of artificial intelligence applications in citizen-state relations.
listelement.badge.access-status Avatud juurdepääs ,
Zero-shot Machine Unlearning using GANs
(Tartu Ülikool, 2024) Ghazal, Ali; El Shawi, Radwa; Jung, Alexander; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
See lõputöö käsitleb masinõppe üliolulist ülesannet, mis hõlmab spetsiifiliste andmete eemaldamist koolitatud masinõppemudelitest, et järgida privaatsusnorme ja parandada andmete kvaliteeti. Tänu tehisintellekti kiirele arengule ja masinõppemudelite laialdasele kasutamisele erinevates rakendustes on tõhusad õppimise eemaldamise meetodid üha pakilisemad. Traditsioonilised lähenemisviisid, nagu mudelite nullist ümberõpe, on suure ressursikulu ja ajapiirangute tõttu ebapraktilised. Lõputöö pakub välja kaks uuenduslikku tehnikat: MuGAN ja zMuGAN. MuGAN on mõeldud stsenaariumide jaoks, millel on piiratud juurdepääs algsele treeningandmete kogumile. See kasutab generatiivseid võistlevaid võrke (GAN), et püüda andmete jaotust mudeli esmase koolituse ajal ja genereerida sünteetilisi andmeid, et sellise päringu saamisel välja õppida. Samamoodi käsitleb zMuGAN olukordi, kus koolitusandmetele pole üldse juurdepääsu. See kasutab GAN-põhist mudeli inversioonitehnikat, et lähendada algset andmekogumit ja hõlbustada õppimist kahjustuste parandamise protsessi kaudu. Mõlemat tehnikat hinnatakse piltide klassifitseerimise ülesannete puhul, eriti klassi unustamisel, rõhutades pildiandmete tundlikku olemust. Pakutud meetodid säilitavad tõhusalt mudeli kasulikkust, tagades samal ajal tõhusa õppimise. Selle lõputöö peamine panus on tugevate ja tõhusate lahenduste pakkumine masinõppeks, mida saab kasutada mudeli arendamise ja juurutamise erinevates etappides. Kasutades GAN-e ja uuenduslikke õppest vabastamise protsesse, pakuvad MuGAN ja zMuGAN selles valdkonnas olulisi edusamme, lahendades pakilise vajaduse praktiliste ja skaleeritavate õppimise eemaldamise tehnikate järele.