DeepFakes Detection and Prevention Using Blockchain

In recent years, the rise of AI and Deep Learning techniques have not only given a completely new perspective to technology, but have also raised many security concerns. Using relatively small amounts of data, Deep Learning Classifiers can create very realistic fake videos (referred to as DeepFakes)

Project Title

DeepFakes Detection and Prevention Using Blockchain

Project Area of Specialization

Blockchain

Project Summary

In recent years, the rise of AI and Deep Learning techniques have not only given a completely new perspective to technology, but have also raised many security concerns. Using relatively small amounts of data, Deep Learning Classifiers can create very realistic fake videos (referred to as DeepFakes) and pose a threat to the integrity of an individual, and an organization as a whole. In the past, these videos have created political uproar and in many instances tainted the reputation of celebrities. To combat this growing threat, we need a tool that provides source of the original content, hence providing the proof of authenticity, therefore ruling out any forged content. In our research, we propose a solution that allows users to trace and track videos back to their original video using Blockchain and Smart Contracts, while also detecting manipulated videos. The framework can be applied to create a platform for the news industry where the authenticity of content is of the utmost importance.

Project Objectives

The objective of our research is to propose a general framework that preserves the integrity of its video content. The framework will allow a viewer to trace back a video even when there exist multiple edited versions of the original video.Viewers can see where any video originated from, who published it, etc. The framework will also be able to detect a Deepfake video, while authentic, unedited videos will have a trust score, so that viewers can distinguish between genuine and fake content. The Blockchain framework also links edited version of videos to their original one. This provides us with an added proof of authenticity.

Project Implementation Method

Our system works for the following use cases:

Use case 1: Video posted by authentic or verified user will get into the system without going through the Deepfake detection API. Its respective hash will be created and stored in the IPFS system.

Use case 2: Video posted by non-verified user will go through the Deepfake detection API. If it is a Deepfake video, it will be detected. Our research focuses on keeping the videos that are not detected as Deepfake with its respective score – that will be calculated using feedback of the viewers.

Deepfake Detection: Deepfake videos will be detected by passing them through an API, which uses machine learning techniques. This API uses algorithms provided in the paper  'Exposing DeepFake Videos By Detecting Face Warping Artifacts' by Yuezun Li and Siwei Lyu. Their method is based on the observations that current DeepFake algorithms can only generate images of limited resolutions, which need to be further warped to match the original faces in the source video. Such transforms leave distinctive artifacts in the resulting DeepFake videos. This method detects such artifacts by comparing the generated face areas and their surrounding regions with a dedicated Convolutional Neural Network (CNN) model.To train the CNN model, we simplify the process by simulating the resolution inconsistency in affine face warpings directly. Specifically, we first detect faces and then extract landmarks to compute the transform matrices to align the faces to a standard configuration. We apply Gaussian blurring to the aligned face, which is then affine warped back to original image using the inverse of the estimated transformation matrix.

Deepfake Prevention: After the video has gone through the detection phase, it can be processed further. If the video is detected as a fake video, it will be added in the system with the verification status as “faked” and reputation score will be set to zero and no computation will be applied to that forged video. But if the video is not detected as a forged video then the video will be added to the system with verified status as “authentic” and reputation score will be calculated gradually on the basis of other user’s/viewer’s feedback. The feedback will be taken by viewers explicitly after the video is uploaded in the system. This will benefit those individual users who have not been verified by the organizations and have posted genuine videos for public benefit. Such videos shall be posted in our system by passing the test of Deepfake detection. Our system does not link the forged videos to the original videos which have already been verified by the system.

Benefits of the Project

Recent advancement in technology of video editing tools, the use of social media sites, and new developments in the fields of artificial intelligence (AI), machine learning (ML), and in particular deep learning (DL) based methods have greatly facilitated the generation of fake content. It is essential to come up with cutting-edge techniques that battle these growing concerns and restore trust in digital media. In the present age, internet users have been left to wonder whether they can ever trust any digital content. Unless there is some mechanism to prove the history and authentication of unadulterated content, the problem only looks to get worse because machine learning techniques will only get better with time. We not only need tools to detect deepfake videos, but also a way to prevent them. Our proposed framework, not only detects a Deepfake video, it also provides authentic, unedited videos will have a trust level, so the users know which videos to trust and which to not. Viewers can see where any video originated from, who published it, etc. Hence, trust in digital content can be restored.

Technical Details of Final Deliverable

Our proposed system will comprise of deepfake detection using Machine learning algorithm and deepfake prevention using Blockchain. For the first part of the research, we work on the detection half of the system, more specifically, extracting frames from videos and recognizing if there are faces in the video and then detecting if those faces have been tampered with. The algorithm displays the fake probability of each video, 1 meaning it is completely fake. High quality videos have clear pixels and so have better detection and frame extraction.

For the second half of the research, we aim to work on deepfake prevention using Blockchain. The results of the deepfake detection will go as an input and will become the parameter as to whether prevention should be done or not. Additionally, at the end of our research, we will propose a system that will cater to videos coming through our system that are not posted by the verified users (Publishers, News Journalists). The genuine content will have its space in the system and will be marked based on this reputation score. The Deepfake detected videos will be given a trust score of zero but will still be available for the viewers and other users with a distinguishing factor.

Final Deliverable of the Project

Software System

Core Industry

IT

Other Industries

Media , Security

Core Technology

Blockchain

Other Technologies

Artificial Intelligence(AI)

Sustainable Development Goals

Industry, Innovation and Infrastructure, Peace and Justice Strong Institutions

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com