Deepfake Detection

EasyChair Preprint 7789

Abstract

Deepfakes are synthetic media that are made by digitally modifying an existing image, video or audio, so that they appear to portray someone else from what they originally did. Deepfakes are popular in spreading malicious false information across the general populace. This is because the quality of the deepfakes being developed is improving with time as a result of breakthroughs in 'Data Science' in general. It has becoming more difficult to distinguish between an original and a deepfake (well-made) media for the same reason. As a result, being able to distinguish between the original and the deepfake becomes critical, as any disinformation spreads like wildfire on social media, causing problems for everyone. The goal of this project was to create a model that, when fed digital media (such as video), could determine whether it was a deepfake or not. The training data consisted of videos that had been pre-processed so that only a few frames from each video were extracted. The retrieved frames are then sent to retinaface, which extracts only the section of the frame (image) that contains a person's face. Utilizing the information gathered in the previous step, the frame is cropped before being subjected to various augmentations and experiments using the XceptionNet and EfficientNet (and its variants) models. To determine the accuracy of the resulting model, a log-loss function was used. The initial model runs resulted in a score of 0.6~0.7, which has since been improved to 0.199. As a result, the project has successfully produced a model that can predict deepfakes (in this case, videos), with an accuracy benchmark of 0.199. This score could be improved even further with additional optimization.

Keyphrases: Deepfake, EfficientNet, Log loss, NLP, RetinaFace, XceptionNet

@booklet{EasyChair:7789,
  author    = {Eshit Bansal and A. Helen Victoria},
  title     = {Deepfake Detection},
  howpublished = {EasyChair Preprint 7789},
  year      = {EasyChair, 2022}}