Detekcija artefaktov generativnih nasprotniških mrež kot pomoč pri detekciji globokih ponaredkov

Križnar, Vid (author), Peer, Peter (mentor), Batagelj, Borut (comentor)

V delu predstavimo nov pristop k detekciji globokih ponaredkov. Globoki ponaredek je tip medija, t.j. slika ali video posnetek, pri katerem je del slike, najpogosteje obraz ali telo digitalno modificirano. Velikokrat so uporabljeni za zle namene, kot je ˇsirjenje dezinformacij; najpogosteje so generirani s pomočjo globokih ali generativnih nasprotniških mrež. Digitalna modifikacija medija pogosto pusti t.i. digitalne artefakte v podatkovnem zapisu medija. Artefakte definiramo kot značilke v podatkih slikovnih elementov na digitalnem mediju, ki nastopijo kot nezaželena posledica modifikacije medija. V delu predstavimo pet metod detekcije globokih ponaredkov s pomočjo detekcije artefaktov generativnih nasprotniških mrež. Predstavljene metode evalviramo na sedmih različnih podatkovnih bazah globokih ponaredkov, ki jih dodatno razdelimo na take, ki so primarno generirane z generativno nasprotniško mrežo, in na te, ki niso. Pokažemo, da predstavljene metode dosegajo obetavne rezultate na pripravljenih podatkovnih bazah.

• Language: Slovenian
• Keywords: globoki ponaredki, generativna nasprotniška mreža, nevronska mreža, artefakt, slikovna biometrija
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-143452
• COBISS.SI-ID: 136512515
• Publication date in RUL: 21.12.2022

KRIŽNAR, Vid, 2022, Detekcija artefaktov generativnih nasprotniških mrež kot pomoč pri detekciji globokih ponaredkov [online]. Master’s thesis. [Accessed 18 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=143452

Secondary language

• Language: English
• Title: Detection of generative adversarial network artefacts as an aid for detecting deepfakes

Abstract

We present a novelty approach to deepfake detection. Deepfake is a type of media, usually a picture or video, in which a part of the picture, most frequently face or body, has been digitally modified. Deepfakes are often used with ill intentions, such as spreading misinformation or opinion formulation. Modification of digital media usually leaves traces, a so-called digital artefacts. Artefacts can be defined as irregularities in digital media which are unwanted consequences of modification. We present five methods for detecting deepfakes by detecting artefacts of generative adversarial networks. We evaluate the presented methods on seven different deepfake databases which are further divided into those that are primarily generated by a generative adversarial network and those that are not. We show that the presented methods achieve promising results on the prepared databases.

• Keywords: deepfakes, generative adversarial network, neural network, artefact, image-based biometrics