TUM Logo

Deepfake Detection Using Sequential Models

Deepfake Detection Using Sequential Models

Supervisor(s): Philip Sperl, Jan-Philipp Schulze
Status: open
Topic: Others
Type of Thesis: Masterthesis
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching

Description

Master Thesis
Deepfake Detection Using Sequential Models
Motivation and Task Description
In recent years, deep neural networks (NNs) have achieved remarkable results and even showed
super-human capabilities in a broad range of domains. Generative Adversarial Networks, a special
class of NNs, synthesize highly realistic images, videos, or audio samples. Recently, this technique
was popularized by so-called “deepfakes”, i.e. manipulated videos with altered faces, gestures, or
actions. Well-made deepfakes may deceive human viewers, potentially discrediting the portrayed
person. This shows the need of automated end-to-end methods to detect deepfakes.
In this thesis, a new approach to detect deepfakes based on our latest findings in adversarial machine
learning is evaluated. For this purpose, state-of-the-art detection tools are further refined to match
the sequential nature of deepfake videos. Based on currently used methods and our concepts, an
end-to-end framework is implemented and tested. Finally, a comparison of current detection systems
is performed and analyzed.
Requirements
Sophisticated programming skills, preferably
in Python
First knowledge of deep learning libraries like
TensorFlow is a plus
Interest in Deep Learning and IT Security
Ability to work self-directed and systematically
Motivation and self-organization
Contact
Philip Sperl Jan-Philipp Schulze
Telefon: +49 89 322-9986-141 Telefon: +49 89 322-9986-195
E-Mail: philip.sperl@aisec.fraunhofer.de E-Mail: jan-philipp.schulze@aisec.fraunhofer.de
Fraunhofer Research Institution for Applied and Integrated Security (AISEC)
Cognitive Security Technologies
Lichtenbergstraße 11, 85748 Garching (near Munich), Germany
https://www.aisec.fraunhofer.de Ausschreibungsdatum: 15. Oktober 2019