# FakeBuster: A DeepFakes Detection Tool for Video Conferencing Scenarios

**Authors:** Vineet Mehta, Parul Gupta, Ramanathan Subramanian, Abhinav Dhall  
**arXiv:** [2101.03321](https://arxiv.org/abs/2101.03321)  
**Date:** 2021-01-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.03321v1/

## Key points

- FakeBuster is a standalone deep learning-based system for detecting manipulated videos during video conferencing and on social media platforms
- It can detect if a person's video has been manipulated or spoofed during a video conference, independent of specific software like Zoom or Skype
- FakeBuster uses a 3D convolutional neural network (CNN) to predict fakeness scores for each segment of a video
- The network is trained using diverse datasets including Deeperforensics, DFDC, VoxCeleb, and locally captured images tailored for video conferencing scenarios
- FakeBuster addresses the concern of DeepFake technology by providing real-time detection of manipulated videos
- It operates independently from specific video conferencing platforms, making it adaptable and versatile
- The use of a 3D CNN ensures accurate prediction of fakeness scores at the segment level
- Users can identify potential impostors or manipulated content during video conferences or on social media platforms
- FakeBuster leverages deep learning techniques and training on diverse datasets to combat DeepFakes effectively
- It ensures authenticity and trustworthiness in video conferencing scenarios and social media interactions.

## Summary

This paper introduces FakeBuster, a novel DeepFake detection tool designed to identify impostors during video conferencing sessions and manipulated faces on social media platforms. Unlike existing solutions, FakeBuster is a standalone deep learning-based system that can detect if a person's video has been manipulated or spoofed during a video conference. It is independent of specific video conferencing software and has been successfully tested with popular applications like Zoom and Skype. FakeBuster utilizes a 3D convolutional neural network (CNN) to predict fakeness scores for each segment of a video. To train the network, the authors combine various datasets including Deeperforensics, DFDC, VoxCeleb, and locally captured images specifically tailored for video conferencing scenarios. This diverse dataset introduces different environments and perturbations, enhancing the generalization capabilities of the deepfake detection network. The proposed FakeBuster tool addresses the growing concern of DeepFake technology by providing an effective solution for detecting manipulated videos in real-time. Its ability to operate independently from specific video conferencing platforms makes it highly adaptable and versatile. The use of a 3D CNN ensures accurate prediction of fakeness scores at the segment level, enabling users to identify potential impostors or manipulated content during video conferences or on social media platforms. Overall, FakeBuster offers an innovative approach to combatting DeepFakes by leveraging deep learning techniques and training on diverse datasets. Its effectiveness in detecting manipulated videos in real-time makes it a valuable tool for ensuring authenticity and trustworthiness in video conferencing scenarios and social media interactions.