DeepFake Video Detection
Date
2020-07-24
Student ID
Research Supervisor
Editor
Journal Title
Volume
Issue
Journal Title
Journal ISSN
Volume Title
Abstract
The rapidly increasing number of photo editing tools like Photoshop and mobile apps such as Snapchat, Adobe Photoshop, PhotoScape, and many more makes it an effortless task to create fake content. We can also argue that the production of counterfeit videos has never been more unaffected. Therefore, our primary goal was to learn the techniques and steps for creating a false video in this case and detect whether an image is fabricated or real (Sultani and Shah, 2019). This paper presents a fundamental learning approach in which computer graphics are separated from authentic photographic images. Our proposed method uses a convolutional neural network with a customized pool to optimize the current algorithms' best-performing feature extraction system. The issue was split into two stages by traditional approaches: visual design or instruction, and prediction. Nevertheless, the research community is now focused on deep learning, inspired by promising results in computer vision. For several forensic applications, such as fake detections and locations, local descriptors based on image noise remains have been incredibly sufficient. The compression which degrades the data significantly does not usually make conventional imaging forensic techniques appropriate for images. A forger can change an image with many different image editing operations when making a forged image. A forensic investigator's interest in developing the forensic algorithms that can detect many operations and manipulations of varying image editing has arisen since each one of them is tested by forensic examiners (Rossler et al., 2019). The rapid progress in the creation and exploitation of digital images has reached a point where significant concerns are raised about society's consequences. At best, this results in a loss of confidence in digital content, but may further cause damage by spreading false or fake information. 8 We suggested an automatic benchmark for identifying facial distortion to standardize estimating detection methods to centralize our project. The parameter is based mainly on the Deep-Fakes, Face2Face, Face Swap, and Neural textures in the random compression and facial manipulation scale (Rossler et al., 2019). We have also created a falsified video detector and an internet browser plugin. The entire scheme fulfills the aim of creating a cyber-safety management system. We also used different neural network architectures to recognize and identify false and real images. We used a c40 compressed video dataset to plan, test, and evaluate the results during our initial stage. Fake images and accurate photos taken from fake and actual videos were used in our data collection. Our model was trained in VGG-19 architecture during the initial stage, with a biased subdivision of data. Later, we equipped our model with various CNN architectures, which provided the best accuracy of 88 percent with Mesonet architecture. During our final training sessions, it was essential to have a vast number of data to manipulate the sequences and create our system. Therefore we decided to take part in a contest hosted by Facebook, DeepFake Detection. It was significant for the constructed model to be impartial and to have regular training in knowing which image is actual and which image is false. Therefore, after we developed the predictive working model, we imported a few other data sets from the competition. In simple terms, 45 of the 50 successful data sets were downloaded. We carried out a detailed study of computer-driven forgery detectors based on the data. We demonstrate that, particularly in the presence of heavy compression, the use of additional domain-specific information increases forgery detection to unparalleled precision, and significantly outperforms human observers.
Description
Keywords
TECHNOLOGY::Information technology::Computer engineering
Citation
Department Name
Electrical and Computer Engineering
Publisher
North South University