Generativne suparničke mreže za izradu neprijateljskih uzoraka

master's thesis

• Title (english): Generative adversarial networks for generating adversarial examples
• Author: Bruna Dujmović
• Mentor: Domagoj Jakobović (mentor)
• Committee member: Domagoj Jakobović (predsjednik povjerenstva)
• Committee member: Marko Đurasević (član povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-06-28, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Suvremeni klasifikacijski modeli, unatoč zavidnim rezultatima na nizu skupova podataka, i dalje znaju biti podložni napadima neprijateljskim uzorcima. Jedan od načina oblikovanja tih ciljano izmijenjenih uzoraka oslanja se na generativne suparničke mreže. Dodavanjem klasifikatora koji se želi napasti u osnovnu GAN arhitekturu, tzv. AdvGAN sposoban je proizvesti uspješne slikovne i zvučne neprijateljske uzorke uvjetovane odgovarajućim originalnim uzorcima. Uz izvornu AdvGAN inačicu, prozvanu image-to-perturbation odnosno audio-to-perturbation AdvGAN-om, razvijena je i noise-to-perturbation inačica koja proizvodi neprijateljske uzorke iz vektora nasumično generiranog šuma. Pokazalo se kako razvijeni AdvGAN-ovi djeluju prilikom izrade uzoraka za MNIST te Speech Commands skupove podataka.

Abstract (english)

Modern classification models, despite the enviable results they achieve on a number of datasets, can still be vulnerable to attacks using adversarial examples. One way of generating these intentionally modified examples relies on generative adversarial networks. By incorporating the targeted classifier into the basic GAN architecture, the so-called AdvGAN is capable of producing successful image and audio adversarial examples conditioned on their unmodified counterparts. In addition to the original AdvGAN version, dubbed "image-to-perturbation" or "audio-to-perturbation" AdvGAN, a "noise-to-perturbation" version has been developed that produces adversarial examples out of random noise vectors. This thesis shows how the developed AdvGANs perform when creating examples for the MNIST and Speech Commands datasets.

• Keywords: neprijateljski uzorci; generativne suparničke mreže; duboko učenje; strojno učenje
• Keywords (english): adversarial examples; generative adversarial networks; deep learning; machine learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:307052
• 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 14:51:29