Logo image
Back
First steps toward detecting and localizing fake contents in videos
Thesis   Open access

First steps toward detecting and localizing fake contents in videos

Tai Duc Nguyen
Master of Science (M.S.), Drexel University
Jun 2021
DOI:
https://doi.org/10.17918/00000439
pdf
Nguyen_Tai_202119.00 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Artificial intelligence--Data processing Biometric identification Pattern perception--Research Common fallacies Cluster analysis Spectral theory (Mathematics) Video recordings Machine Learning
Misinformation, especially in multimedia contents such as images and videos, has become ubiquitous due to the availability of computer software such as Adobe Photoshop and deep neural network techniques that generate deepfakes. As the amount of fake content increases and its quality improves, it is crucial that research must be done to combat this issue. Recent works have delivered promising results on images, especially those which utilized deep learning to extract hidden forensic traces left by content manipulation and falsification. Research into video, however, is much further behind. Currently, there are no methods to detect and localize fake or manipulated content that have been explicitly designed for video. This thesis represents the initial effort of building such method by addressing three fundamental problems: 1) forensic traces of video frames are different from images, 2) in modern video there are three different types of frames, each with different traces, and 3) the computational complexity of analyzing videos is much higher compared to images. In order to meet these challenges, we first propose a novel system for detecting and localizing fake content in videos by adapting existing work on Forensic Similarity Graphs. Next, we present an approach to address computational complexity issue associated with video. Finally, we create new datasets and use them to evaluate the performance of our system and other approaches.

Metrics

89 File views/ downloads
99 Record Views

Details

Logo image