Description
Installation
Quick Start
Maintainers
Deep-PoC is a deepFake detection tool designed to detect deepfakes from videos or images using artificial intelligence.
The detection relies on two main weakness of deepfake : The mouth and the eyes. Deep-PoC focuses on these main parts to make its prediction.
The project is comprised of 2 parts : The Web-App, and AI.
The website can be found at https://deepoc.poc-innovation.com/
The project's AI was created using the library called PyTorch. The AI is a imple CNN with 4 layers of convolution and 3 linear layers. The prediction is between 0 and 1, (0 being the detction of a deepfake and 1 being the detection of a real face).
The Ai suffers from a lack of diversity in the dataset, most of the deepfake come from https://thispersondoesnotexist.com, therefore it lacks the high capacity to detect deepfakes generated differently.
The web-app was created with django with the help of the dropzonejs library (https://www.dropzonejs.com/).
It is comprised of a simple drag and drop feature, to add the video of your liking to be analyzed by the AI.
Here is the list of supported extension:
Extension | Operational |
---|---|
.mp4 | ✔️ |
.jpeg | ✔️ |
.jpg | ✔️ |
.png | ❌ |
You'll need Python3 or higher, pip3 and docker installed. Install the requirements with pip3 install -r requirements.txt
git clone git@github.com:PoCInnovation/Deep-PoC.git
cd Deep-PoC/DeepPoc
python manage.py runserver
docker-compose up
Different scripts can be used to create / update and train your own dataset and AI. These scripts are located in the: ./src/scripts/
directory.
They are to be launched from the root of the project, here is an exemple:
python ./src/scripts/manual_test.py -h
Each script possesses a (-h or --help)
option to view the usage of the script.
The Dataset used during this project doesn't have an easy source (For now).
To build your own dataset, you have a script to extract Deepfakes from https://thispersondoesnotexist.com.
This only goes for deepfake, to get real images, I suggest the following dataset: https://github.com/NVlabs/ffhq-dataset.