Tasneem Wael, Adel Ahmed, Muhammed Adel
Supervised by: Dr. Taraggy Ghanim, Eng. Youssef Talaat
7 February 2022
The usage of machine learning and deep learning in creating forged videos has increased throughout the past few years. New methods like GANs are used to make the fake videos look more realistic which makes it more difficult to tell the difference between the real and the altered media. Our system works on multiple CNN classifiers in an ensemble to detect the forged videos and images as it will help bring back the safety and reliability of using social media.
1.1 Purpose of this document
Deepfake is a type of synthetic media in which a person’s face is replaced with that of someone else in an existing photograph or video. Deepfake technology dates back to the nineteenth century and has been evolving ever since. Its algorithm uses artificial intelligence and machine learning to modify visual and audio content. Deepfake entails the training of generative neural networks such as “autoencoders” and “GANs” (generative adversarial networks). Blackmail, pornography, politics, art, internet comics, acting, social networking, and socket puppets are all employed by Deepfake. Deepfake technology is constantly being improved, allowing for the development of offrauds, fraud credibility, and authenticity. Several social media networks have already taken the initiative and built technology to detect deep fake content.
1.2 Scope of this document
The proposed system will be able to train on deepfake videos from datasets like:
STYLEGANS2 using RESNET or Siamese network. The classifiers then will be used to detect different deepfake methods such as(face2face, faceswap, deepfake, and neural textures). Our system will test and train on different datasets to avoid any over-fitting.
1.3 System Overview
DeepFake detection is necessary following all of the videos that have appeared in recent years. Deepfake technology has evolved and is becoming more difficult to address as time goes on, causing future problems. The project will attempt to evaluate videos across several platforms in order to avoid the spread of incorrect information. The proposed method will take a fake video generated by GANs from a deep fake video dataset and train it with resnet or a siamese network. The classifier will then be used to test a variety of deepfake technologies, including face2face, faceswap, neural textures, and deepfake. To avoid overfitting, the training dataset will not be used in the testing dataset, and testing will be performed on multiple datasets (Faceforensics ++, DFDC, and GANS). TheConvolution Neural Network technique will be used in this project. The system will be able to extract the victim’s facial traits and compare them to the original subjects to look for discrepancies.
1.4 System Scope
The proposed framework will center on numerous aspects:
- The framework will work on combined datasets; meaning that it’ll incorporate more than one dataset comprising of fake and genuine recordings.
- The framework will test on low quality and high quality video recordings.