Napadi na neuronske mreže za detekciju objekata umetanjem stražnjih vrata

master's thesis

• Title (english): Backdoor Attacks Upon Object Detection Neural Networks
• Author: Tomislav Prhat
• 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: 2023-07-04, Croatia

Abstract

U ovom radu dan je opis glavnih dijelova i principa rada modela za detekciju te prikriveni napadi na takve modele. Kroz poglavlja su se postepeno objašnjavali dijelovi, počevši s opisom neuronskih mreža kao temeljnog bloka strojnog učenja, pregleda srodnih radova, detaljnog objašnjenja prikrivenih napada i raznih strategija napada, zatim su objašnjene vrste okidača korištene u ovom radu, te konačno implementacija prikrivenih napada umetanjem stražnjih vrata u modele za detekciju objekata. Eksperimenti su provedeni kako bi se ispitala učinkovitost i rezultati prikrivenih napada na detekciju objekata. Rezultati eksperimenata pružaju uvid u utjecaj i rizike koje prikriveni napadi mogu imati na performanse neuronskih mreža u detekciji objekata. Konačni rezultati ovog rada su zadovoljavajući, dvije vrste okidača su imale lošije rezultate, dok su dvije vrste imale bolje rezultate nego u polaznom radu.

Abstract (english)

This paper describes the main parts and working principles of detection models and backdoor attacks on such models. Through the chapters, the parts were gradually explained, starting with the description of neural networks as a fundamental block of machine learning, an overview of related works, detailed explanations of stealth attacks and various attack strategies, then the types of triggers used in this work were explained, and the final implementation of backdoor attacks by inserting back doors in object detection models. Experiments were conducted to test the effectiveness and results of backdoor attacks on object detection. The results of the experiments provide insight into the impact and risks that backdoor attacks can have on the performance of neural networks in object detection. The final results of this work are satisfactory, two types of triggers had worse results, while two types had better results than in the initial paper.

• Keywords: Strojno učenje; sigurnost; napadi na neuronske mreže; neuronske mreže; prikriveni napadi
• Keywords (english): Machine Learning; Security; Neural Network Attacks; Neural Networks; Backdoor Attack
• Language: croatian
• URN:NBN: urn:nbn:hr:168:576093
• 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 10:13:33