Anonimizacija podataka u postupcima strojnog učenja

master's thesis

• Title (english): Data anonymization in machine learning methods
• Author: Patricia Nadarević
• Mentor: Boris Vrdoljak (mentor)
• Committee member: Boris Vrdoljak (predsjednik povjerenstva)
• Committee member: Dalibor Krleža (član povjerenstva)
• Committee member: Mirta Baranović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

S povećanjem količine dostupnih podataka i napredovanjem računala, primjena strojnog učenja doživjela je nevjerojatan rast u akademskim istraživanjima i u stvarnom životu, a performanse modela strojnog učenja uvelike ovise upravo o podacima nad kojima model uči. Glavni problem anonimizacije u kontekstu strojnog učenja je pronaći balans između korisnosti i sigurnosti – kako zaštititi podatke, a da se istovremeno sačuva dovoljna količina korisnih informacija na temelju kojih model može učiti.
U sklopu ovog diplomskog rada provedeni su postupci analize i obrade nad dva skupa podataka – Melbourne Housing Market i Adult Income, koristeći Python programski jezik, nakon čega su izgrađeni modeli linearne regresije za određivanje cijene nekretnine, klasifikacijski model za određivanje godišnjih primanja te model LSTM mreže. Također su istraženi različiti pristupi anonimizacije podataka koji su potom i primijenjeni nad korištenim skupovima podataka. Primijenjene su Python biblioteke za k-anonimizaciju, kojima su podaci anonimizirani metodama generalizacije, maskiranja i supresije. Primijenjena je i metoda generiranja sintetičkih podataka. Naposljetku, nad anonimiziranim podacima ponovno su izgrađeni modeli strojnog i dubokog učenja, kako bi se vidio utjecaj anonimizacije na točnost modela.

Abstract (english)

With the increase in the amount of data available and the advancement of computers, the application of machine learning has experienced incredible growth in academic research and in real life, and the performance of the models largely depend on the training data. The main problem of anonymization in the context of machine learning is finding a balance between utility and security - how to protect data while preserving a sufficient amount of useful information from which the model can learn.
As part of this thesis data analysis and processing procedures were performed on two datasets – Melbourne Housing Market and Adult Income, using Python programming language, after which models of linear regression for determining the price of real estate, classification model for determining annual income and LSTM network model were built. Different approaches to data anonymization were investigated, which were then applied to the datasets used. Python libraries used for k-anonymization, in which data was anonymized using methods of generalization, masking and suppression. The method of generating synthetic data was also applied. Finally, machine and deep learning models were rebuilt over anonymized data, to see the impact of anonymization on model accuracy.

• Keywords: osobni podaci; anonimizacija podataka; strojno učenje; duboko učenje; linearna regresija; klasifikacija
• Keywords (english): personal information; data anonymization; machine learning; deep learning; linear regression; classification
• Language: croatian
• URN:NBN: urn:nbn:hr:168:436754
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 13:18:29