Suparnički napadi u obradi prirodnog jezika

master's thesis

• Title (english): Adversarial Attacks in Natural Language Processing
• Author: Josip Jukić
• Mentor: Domagoj Jakobović (mentor)
• Committee member: Domagoj Jakobović (predsjednik povjerenstva)
• Committee member: Marko Đurasević (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Motivacija za problem izgradnje suparničkih primjera, opisana je u uvodnom poglavlju. Obrada prirodnog jezika i koncepti suparničkih napada ukratko su opisani. Glavni cilj rada je razvoj raznolikih napada u području prirodnog jezika. Naglasak je stavljen na učinkovite napade i na metode koje mogu očuvati strukturalnu i semantičku cjelovitost teksta. Razvijeni su novi pristupi napada. Oslanjaju se na procjenu značajnosti pojedinih riječi u tekstu, kao i na intenzivno korištenje modela za leksičku zamjenu te ručno rađenih reprezentacija riječi. Provedeni su brojni eksperimenti kako bi se procijenile i usporedile izvedbe pojedinih suparničkih napada. Rad je zaključen kratkih pregledom mogućih obrambenih mehanizama i uvidima koji su stečeni na temelju dobivenih rezultata.

Abstract (english)

Motivation for this assignment, crafting adversarial examples for textual machine learning models, is presented in the introductory chapter. Natural language processing and adversarial attack concepts, important for understanding the experiments, are briefly described. Main topic of this assignment is development of various attacks in the natural language domain. Emphasis is put on effective attacks and methods that can preserve structural and semantic integrity of text. Novel approaches have been designed. They rely on evaluating importance of certain words in text, along with the heavy use of lexical substitution models and custom-crafted word embeddings. A series of experiments had been conducted in order to evaluate and compare different adversarial attacks. The thesis is concluded with short overview of possible defence mechanisms and insights gathered from numerous results.

• Keywords: suparnički; napad; tekst; prirodni jezik; klasifikacija; zaključivanje; strojno učenje; duboko učenje
• Keywords (english): adversarial; attack; text; natural language; classification; entailment; machine learning; deep learning
• Language: english
• URN:NBN: urn:nbn:hr:168:954663
• 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 12:32:57