prikaz prve stranice dokumenta Otkrivanje privatnih informacija u modelima strojnog učenja
No public access
master's thesis
Otkrivanje privatnih informacija u modelima strojnog učenja
Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing, 2022. urn:nbn:hr:168:355946
Škudar, Silvije
University of Zagreb
Faculty of Electrical Engineering and Computing

Institutional repository: FER Repository

Cite this document

Škudar, S. (2022). Otkrivanje privatnih informacija u modelima strojnog učenja (Master's thesis). Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing. Retrieved from https://urn.nsk.hr/urn:nbn:hr:168:355946

Škudar, Silvije. "Otkrivanje privatnih informacija u modelima strojnog učenja." Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2022. https://urn.nsk.hr/urn:nbn:hr:168:355946

Škudar, Silvije. "Otkrivanje privatnih informacija u modelima strojnog učenja." Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2022. https://urn.nsk.hr/urn:nbn:hr:168:355946

Škudar, S. (2022). 'Otkrivanje privatnih informacija u modelima strojnog učenja', Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, accessed 14 January 2025, https://urn.nsk.hr/urn:nbn:hr:168:355946

Škudar S. Otkrivanje privatnih informacija u modelima strojnog učenja [Master's thesis]. Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing; 2022 [cited 2025 January 14] Available at: https://urn.nsk.hr/urn:nbn:hr:168:355946

S. Škudar, "Otkrivanje privatnih informacija u modelima strojnog učenja", Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, 2022. Available at: https://urn.nsk.hr/urn:nbn:hr:168:355946

Metadata
TitleOtkrivanje privatnih informacija u modelima strojnog učenja
Title (english)Leakage of private information in machine learning models
AuthorSilvije Škudar
MentorAnte Đerek (mentor)
Committee memberAnte Đerek (predsjednik povjerenstva)
Committee memberMarin Golub (član povjerenstva)
Committee memberMarko Đurasević (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb
Defense date and country2022-07-07, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Abstract
Nedavna istraživanja su pokazala da modeli strojnog učenja imaju sposobnost pamćenja osjetljivih podataka, čime dolazi do mogućnosti napada utvrđivanjem pripadnosti. Maliciozni korisnik koji modelu pristupa kao crnoj kutiji može iskoristiti taj napad i s visokom vjerojatnošću zaključiti je li određeni podatak bio korišten u svrhu treniranja modela, čime dolazi do narušavanja povjerljivosti tog podatka. Zadatak ovog diplomskog rada bio je istražiti moguće napade na privatnost u modelima strojnog učenja kao i njihovu prevenciju i okolnosti koje ih omogućuju. U tu svrhu, koristili smo postojeće metode napada utvrđivanjem pripadnosti i metrike, i reproducirali nekoliko eksperimenata kako bismo promatrali utjecaj prenaučenosti, arhitekture modela, optimizacijskih metoda i sastava skupa podataka za treniranje na uspješnost tih napada. Nadalje, također smo pokušali korelirati značajke podataka i rizik privatnosti pomoću analize glavnih komponenti. Konačno, eksperimentirali smo s iscrtavanjem decizijskih granica i isprobali smo novu metodu za sintezu napadačkih podataka koja se temelji na evolucijskoj strategiji CMA-ES.
Abstract (english)
Recent research has shown that machine learning models have the capacity to memorise sensitive data, opening up an avenue for the membership inference attack. Given a black- box API, a malicious user can perform the attack and learn, with high confidence, whether a certain data record was used for training the model, which constitutes a confidentiality breach. The objective of this thesis was to investigate the potential privacy attacks on machine learning models as well as the circumstances which lead to privacy leakage and its prevention. To this end, we make use of the existing membership inference attacks and metrics and reproduce a number of experiments to observe the impact of overfitting, model architecture, optimisation methods and training data composition on the success of the attacks. Furthermore, we also attempt to correlate the data features to privacy risk using principal component analysis. Finally, we also experiment with plotting the decision boundaries and we try a new shadow data synthesis method using CMA-ES.
Keywords
Keywords (english)
Languageenglish
URN:NBNurn:nbn:hr:168:355946
Study programmeTitle: Computing
Study programme type: university
Study level: graduate
Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsClosed access
Terms of useLicence In copyright icon
Public note
Created on2023-01-13 13:57:23