Napadi na neuronske mreže korištenjem dubokih prikrivenih značajki

master's thesis

• Title (english): Deep Hidden Feature Backdoor Attacks Upon Neural Networks
• Author: Mateo Stanić
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-03, Croatia

Abstract

U ovom radu razmatramo napad umetanja stražnjih vrata u modelima dubokog učenja. Cilj napada je da za „otrovane“ primjerke model daje ciljnu klasu dok za čiste uzorke funkcionira kao normalan klasifikacijski model. Popratni cilj napada u ovom radu je također ostvariti kvalitetan okidač za stvaranje otrovanih primjeraka te korištenjem naučenih značajki međuslojeva ostvariti veću „nevidljivost“ od standardnih napada. Opisana je teorijska podloga te implementacija u programskim okruženjima. Napadi su testirani na dva modela i dva poznata skupa podataka i uspoređeni s ostalim vrstama napada. Razmotrili smo ograničenja napada i implementacije te donijeli zaključke o uspješnosti izvedenog napada.

Abstract (english)

In this paper, we consider backdoor attacks in deep learning models. The aim of the attack is that for "poisoned" samples the model predicts the target class, while for clean samples it functions as a normal classification model. The goal of the attack in this work is also to achieve a high-quality trigger for the creation of poisoned samples and to achieve greater "invisibility" than standard attacks by using the learned features of the intermediate layers. The theoretical background along with its implementation in programming environments is described in detail. The attacks were tested on two models and two known datasets and compared with other types of attacks. We considered the limitations of the attack and its implementation, and reached conclusions about the success of the attack.

• Keywords: duboko učenje; trovanje; stražnja vrata
• Keywords (english): deep learning; poisoning; backdoor
• Language: croatian
• URN:NBN: urn:nbn:hr:168:370421
• 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: 2023-10-11 10:30:55