Napadi na modele za detekciju objekata zasnovani na trovanju podataka

master's thesis

• Title (english): Attacks on object detection models based on data poisoning
• Author: Ivan Sabolić
• Mentor: Marko Čupić (mentor)
• Committee member: Marko Čupić (predsjednik povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-12, Croatia

Abstract

Detekcija objekata u slikama važan je zadatak računalnog vida s mnogim zanimljivim primjenama. U posljednje
vrijeme vrlo zanimljive rezultate postižu duboki modeli zasnovani na konvolucijama i slojevima pažnje. Međutim,
veliki kapacitet tih modela omogućava sofisticirane trojanske napade zasnovane na trovanju podataka za
učenje. U okviru ovog rada, predstavljene su osnovne postavke trojanskih napada.
Uhodani su jednostavni napadi na model za detekciju objekata te su prikazani rezultati takvih napada.
Ispitan je utjecaj finog ugađanja na napade trovanjem podataka u svrhu obrane i uspješnijeg napada.
Također je predložena metoda obrane od napada trovanjem podataka zasnovana na procjeni izglednosti
značajki dobivenih samonadziranim učenjem. Provedeni su eksperimenti koji pokazuju uspješnost navedene
obrane.

Abstract (english)

Object detection in images is an important computer vision task with many interesting applications.
Very interesting results were recently achieved by deep models based on convolutions and attention. However,
the large capacity of these models enables trojan attacks based on data poisoning.
This paper presents the basic settings of the Trojan attack.
Simple attacks on the object detection model have been implemented and the results of such attacks are presented.
The influence of fine-tuning on the data poisoning attack was examined for the purpose of defense and more successful attacks.
Another defense method against data poisoning attacks based on likelihood estimation
of features obtained by self-supervised learning is also proposed.
Experiments were carried out that show the success of the mentioned
defense.

• Keywords: trojanski napadi; detekcija objekata; napadi trovanjem podataka; obrana od napada trovanjem podataka
• Keywords (english): trojan attacks; object detection; data poisoning attacks; defenses from data poisoning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:015296
• 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 10:23:14