magistrsko delo
Anže Mur (Author), Peter Peer (Mentor), Borut Batagelj (Co-mentor)

Abstract

Hiter razvoj na področju globokega učenja je privedel do množične adaptacije koncepta globokih ponaredkov. Globoki ponaredki so sintetični mediji, ki so velikokrat ustvarjeni zlonamerno in tako postajajo vse večji izziv za sodobno družbo. S tem razlogom je ključnega pomena razviti robustne in učinkovite metode za zaznavanje globokih ponaredkov, da bi preprečili njihovo zlonamerno uporabo. V našem delu smo implementirali konvolucijsko nevronsko mrežo Xception, ki smo jo nadgradili z arhitekturo, da deluje po principu dveh ločenih vej učenja. Prva veja se uči samo na manipulirani obrazni regiji, druga veja pa si za napovedovanje končnega rezultata pomaga tudi z nespremenjenimi regijami izven obraza. Implementirana dvovejna arhitektura izboljša delovanje osnovnega Xception modela za 2,45 % vrednosti AUC iz prvotne vrednosti 69,76 %. Implementirana modela smo dodatno učili tudi na novo ustvarjeni sintetični podatkovni zbirki artefaktov globokih ponaredkov, kjer Xception model doseže izboljšavo vrednosti AUC osnovnega modela za 12,5 % iz prvotne vrednosti 69,76 %.

Keywords

globoki ponaredki;konvolucijske nevronske mreže;detekcija globokih ponaredkov;slike;mediji;računalništvo in informatika;magisteriji;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [A. Mur]
UDC: 004.85:7.061(043.2)
COBISS: 170117635 Link will open in a new window
Views: 171
Downloads: 40
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Deepfake detection based on the analysis of the manipulated and non-manipulated regions
Secondary abstract: The rapid development in the field of deep learning has led to the general adoption of the deepfake concept. Deepfakes are synthetic media that are often created maliciously and therefore pose an increasingly significant challenge to modern society. For this reason, it is crucial to develop robust and effective methods for detecting deepfakes to prevent their malicious use. In our work, we implemented the Xception convolutional neural network, which we upgraded with an architecture that operates on the principle of two separate learning branches. The first branch learns only on the manipulated facial region, while the second branch uses non-manipulated regions outside of the face region to predict the final result. The implemented dual-branch architecture improves the performance of the baseline Xception model by 2.45 % in terms of AUC value from the original value of 69.76 %. We additionally trained the implemented models on a newly created synthetic dataset of deepfake artifacts, where the Xception model achieves a 12.5 % improvement in the AUC value of the baseline model with the value of 69.76 %
Secondary keywords: deep learning;convolutional neural networks;deepfake detection;pictures;media;computer science;computer and information science;master's degree;Globoko učenje (strojno učenje);Nevronske mreže (računalništvo);Ponarejanje in ponaredki;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Master's thesis/paper
Study programme: 1000471
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 59 str.
ID: 19963981