DeepFake Detection on Publicly Available Datasets using Modified AlexNet
Daniel Xie, Prosenjit Chatterjee, Zhipeng Liu, Kaushik Roy, Kossi Edoh
Abstract
Deep learning has been applied successfully in many areas, including computer vision, natural language processing, cyber physical systems and big data analytics. Recently, a synthesis of deep leaning techniques has been deployed to create fake images and videos that are not easily distinguishable from the real ones; this technology is known as DeepFakes. In this paper, we looked at various DeepFakes related datasets and created a model in order to identify whether a frame of a video is fake or real. This is important as videos can be easily manipulated in a way that can spread misinformation, and that can cause major problems in the world today, especially if the videos have political implications. In order to create a model, we utilize a modified AlexNet constructed of an arrangement of 6 layers: convolution2d, max pooling, dense, flatten, activation and dropout layers. UADFV, FaceForensics++, and Celeb-DF are the 3 datasets used in this research. There are many publicly available datasets, however, we found the UADFV, FaceForensics++ and Celeb-DF to be the most convenient in how the data was formatted. All data for each dataset was organized into videos of varying classes. While the UADFV dataset is straightforward and only has 2 classes: real and fake, the FaceForensics++ dataset looks at the various kinds of video manipulations and has 5 different classes. Our model was able to achieve a 9S.73% accuracy when identifying whether a video is real or fake on the UADFV dataset, a slightly adjusted version was able to accomplish an 87.49% accuracy with the FaceForenisics++dataset, and reached 98.85% accuracy on the Celeb-df dataset.