kth.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Accuracy and Robustness of State of the Art Deepfake Detection Models
KTH, School of Electrical Engineering and Computer Science (EECS).
KTH, School of Electrical Engineering and Computer Science (EECS).
2023 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE creditsStudent thesisAlternative title
Precision och Robusthet hos Bästa Tillgängliga Deepfake-detektionsmodeller (Swedish)
Abstract [en]

With the evolution of artificial intelligence a lot of people have started getting worried about the potential dangers of deepfake images and videos, such as spreading fake videos of influential people. Several solutions to this problem have been proposed with some of the most efficient being convolutional neural networks for face detection in order to differentiate real images from deepfake images generated with a generative adversarial network. One of the currently most prevalent models is the VGGFace which is further analyzed in the report. This project explores how different hyperparameters affect the effectiveness of existing convolutional neuralnetworks aswell as the robustness in the models. The hyper-parameter that had the biggest effect on accuracy was the amount of conovultion layers in each step of the network. The results showed that while deepfake detection models showed high accuracy on the test set, they are lackluster when it comes to the robustness. The models showed a clear sensitivity for the resolution of test images. This is an issue that can be solved through resizing, this report shows the more concerning issue where the model had a 47 percentage point reduction in accuracy when tested on a different dataset that had fake images generated with a different generative adversarial network. The main takeaways from the project is that current deepfake detection models have to work on generalization in order to effectively classify images.

Abstract [sv]

Med utvecklingen av artificiell intelligens har många människor börjat bli oroliga över de potentiella farorna med deepfake-bilder och -videor, exempelvis spridning av falska videor på inflytelserika människor. Flera lösningar på detta problem har föreslagits, varav några av de mest effektiva är konvolutionsneurala nätverk för ansiktsdetektion för att kunna skilja på verkliga bilder och deepfake-bilder som genererats med hjälp av ett generativa motståndarnätverk. En av de främsta nuvarande modellerna kallas VGGFace och analyseras vidare i rapporten. Projektet utforskar hur olika hyperparametrar påverkar effektiviteten hos befintliga konvolutionsneurala nätverk samt undersöker hur robusta modellerna är. Hyperparametern som hade störst effekt på noggrannheten var antalet konvolutionslager i varje steg av nätverket. Resultaten visade att trots deepfake-detektionsmodellernas höga noggrannhet på testdatan så var de bristfälliga gällande robustheten. Modellerna visade tydlig känslighet för upplösningen på testbilder. Detta är ett problem som kan lösas genom att ändra upplösningen på bilderna, men rapporten visar däremot ett mer oroväckande problem där modellen hade en minskning på 47 procentenheter i noggrannhet när den testades på ett annat dataset som hade deepfake-bilder genererade med ett annat generativt motståndarnätverk. De huvudsakliga slutsatserna från projektet är att nuvarande deepfake-detektionsmodeller måste arbeta med generalisering för att kunna klassificera bilder på ett effektivt sätt.

Place, publisher, year, edition, pages
2023. , p. 15
Series
TRITA-EECS-EX ; 2023:285
National Category
Computer Sciences
Identifiers
URN: urn:nbn:se:kth:diva-335982OAI: oai:DiVA.org:kth-335982DiVA, id: diva2:1795907
Subject / course
Computer Science
Educational program
Master of Science in Engineering - Computer Science and Technology
Supervisors
Examiners
Available from: 2023-09-12 Created: 2023-09-11 Last updated: 2023-09-12Bibliographically approved

Open Access in DiVA

fulltext(842 kB)1469 downloads
File information
File name FULLTEXT01.pdfFile size 842 kBChecksum SHA-512
84e4a4a8248dcc49d077df47caba92ca994fb4b3ac746a55c428cc9977080c63440b5b41e499791202c262f6ad17907239c81e0b6977d45472ac9436d2543b5f
Type fulltextMimetype application/pdf

By organisation
School of Electrical Engineering and Computer Science (EECS)
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 1470 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 571 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf