ETD

Digital archive of theses discussed at the University of Pisa

 

Thesis etd-06242021-124424


Thesis type
Tesi di laurea magistrale
Author
COCCOMINI, DAVIDE ALESSANDRO
URN
etd-06242021-124424
Thesis title
Design and Development of Transformer-based Methods for Video Deepfake Detection
Department
INGEGNERIA DELL'INFORMAZIONE
Course of study
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Supervisors
relatore Prof. Falchi, Fabrizio
relatore Prof. Gennaro, Claudio
relatore Dott. Messina, Nicola
Keywords
  • Anomaly Detection
  • Deepfakes
  • Computer Vision
  • Deep Learning
  • TimeSformers
  • Vision Transformers
  • Transformers
Graduation session start date
23/07/2021
Availability
Full
Summary
La tesi si pone come obiettivo di affrontare il problema della video deepfake detection mediante tecniche innovative di deep learning basate sui Transformers, investigandone l'efficacia. A tale scopo è stata prima effettuata un'analisi degli approcci precedentemente utilizzati in questo settore e delle varie architetture già presenti in letteratura identificandone quelle più adatte ed effettuando alcuni test preliminari su task di anomaly detection. Successivamente sono stati progettati e sviluppati diversi modelli ibridi basati su EfficientNet e vari tipi di Vision Transformers, riuscendo a raggiungere risultati vicini allo stato dell'arte.

This thesis aims to tackle the problem of video deepfake detection by means of innovative deep learning techniques based on Transformers, investigating their effectiveness. To this end, an analysis of the approaches previously used in this sector and of the various architectures already present in the literature was carried out, identifying the most suitable ones and carrying out some preliminary tests on anomaly detection tasks. Subsequently, various hybrid models based on EfficientNet and various types of Vision Transformers were designed and developed, achieving results close to the state of the art.
File