Napadi na neuronske mreže umetanjem stražnjih vrata

master's thesis

• Title (english): Backdoor attacks on deep neural networks
• Author: Jura Milić
• 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: 2024-07-15, Croatia

Abstract

U posljednjim godinama, umjetna inteligencija napravila je velike korake u raznim domenama uključujući i detekciju objekata, što je dovelo do ključnih dijelova u svakodnevnim slučajevima poput samovozećih automobila i sigurnosnih sustava. No, robustnost i sigurnost tih modela su postali veliki razlog za brigu, pogotovo s dočekom sofisticiranih vektora napada poput backdoor napada. Ovaj rad će istražiti implementaciju backdoor napada na modele detekcije objekata, s naglaskom na Faster R-CNN modele, koji su vrlo iskorišteni u tom polju. Uspoređujemo trenutnu situaciju u umjetnoj inteligenciji s već dobro razvijenom sigurnosti u kriptografiji, čime naglašavamo potrebu za rigoroznim obranama protiv takvih napada.

Istraživanje će ući u detalje četiri tipa backdoor napada: OGA (napad dodavanja objekta), RMA (napad krive klasifikacije unutar regije), GMA (napad krive klasifikacije na globalnoj razini), te ODA (napad micanja objekta). Generiranjem otrovanih skupova podataka i treniranjem Faster R-CNN modela na tim skupovima podataka, analizirat ćemo utjecaj tih napada. Učinkovitost tih napada je evauluirana korištenjem metrika poput srednje prosječne preciznosti (mAP) i stopom uspješnosti napada (ASR), čime naglašavamo osjetljivost modela detekcije objekata na backdoor napade.

Abstract (english)

In the last few years, artificial intelligence has made great strides in various domains including object detection which led to key components in every-day applications like self driving cars and security systems. But, the robustness and security of these models have become a cause for concern, especially with the arrival of sophisticated attack vectors like backdoor attacks. In this work, we examine an implementation of a backdoor attack on object detection models with emphasis on Faster R-CNN models which have been thoroughly used and researched in the field. We compare the current state of artificial intelligence with cryptography, emphasizing the need for rigorous defenses against such attacks.

The research of this work will go into details on four different types of backdoor attacks: OGA (object generation attack), RMA (regional misclassification attack), GMA (global misclassification attack), and ODA (object dissapearance attack). Generating posioned datasets and training a Faster R-CNN model on these datasets, we will analyze the impact of such attacks. The efficiency of these attacks will be evaluated using metrics such as mAP (mean average precision) and ASR (attack success rate). These metrics will emphasize the vulnerability of object detection models to backdoor attacks.

• Keywords: detekcija objekata; backdoor napadi; računalni vid; duboko učenje; Faster R-CNN model; umjetna inteligencija
• Keywords (english): object detection; backdoor attacks; computer vision; deep learning; Faster R-CNN model; artificial intelligence
• Language: croatian
• URN:NBN: urn:nbn:hr:168:322075
• 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: Open access
• Created on: 2024-10-12 21:37:21