DeepSafe 🐶 - Open Source DeepFake Detection

DeepSafe is a Streamlit-based web application for DeepFake detection, offering an easy-to-use interface for analyzing images and videos. Users can add their own deepfake detection models and compare them with existing models out of the box.

WebApp

Live here (Limited access, for full access please contact me).

Table of Contents

• Features
• Usage
• Installation
• Additional Sections
• WebApp
• Future Work
• Contributing
• Acknowledgments
• License
• Contact Information

Features

✨ Multi-model Support: Users can select from multiple DeepFake detection models for both images and videos.
📁 File Upload: Supports uploading images (jpg, png, jpeg) and videos (mp4, mov).
🌐 URL Input: Allows users to input URLs for image or video analysis.
⚙️ Processing Unit Selection: Option to use GPU for supported models (default is CPU).
📊 Result Visualization:

• Displays DeepFake detection stats in a table format.
• Provides downloadable CSV of detection results.
• Visualizes results with bar charts for DeepFake probability and inference time.

Usage

1. Select the "Detector" option from the sidebar.
2. Upload an image/video or provide a URL.
3. Choose the DeepFake detection model(s) you want to use.
4. Optionally select GPU processing if available.
5. Click "Real or Fake? 🤔" to start the analysis.
6. View the results in the displayed charts and tables.

Additional Sections

• Examples: View sample DeepFakes by selecting "Examples" from the sidebar.
• About: Learn about the detectors used in the app and their original authors.
• Learn: Access educational resources about DeepFakes.

Installation

1. Clone the repository:

   git clone https://github.com/siddharthksah/DeepSafe
   cd DeepSafe

2. Create a conda environment:

   conda create -n deepsafe python==3.8 -y
   conda activate deepsafe

3. Install dependencies:

   pip install -r requirements.txt

4. Download Model Weights from Google Drive:

   

   from pydrive.auth import GoogleAuth
   from pydrive.drive import GoogleDrive
   import os
   
   # Authenticate and create the PyDrive client.
   gauth = GoogleAuth()
   gauth.LocalWebserverAuth()
   drive = GoogleDrive(gauth)
   
   # Specify the folder ID (the part after 'folders/' in the URL)
   folder_id = '1UmMTuXPmu-eYfskbrGgZ1uNXceznPQ6o'
   
   # Create 'models' directory if it doesn't exist
   if not os.path.exists('models'):
       os.makedirs('models')
   
   # List all files in the folder
   file_list = drive.ListFile({'q': f"'{folder_id}' in parents and trashed=false"}).GetList()
   
   # Download each file to the 'models' directory
   for file in file_list:
       file.GetContentFile(os.path.join('models', file['title']))
   
   print("Download complete.")

5. Start the application:

   streamlit run main.py

Benchmarking

DeepSafe includes a powerful benchmarking feature that allows users to benchmark their datasets against selected deepfake detection models. The results include accuracy, precision, and recall metrics, along with detailed visualizations.

Benchmark Your Datasets

1. Select Dataset Type: Choose between Image or Video datasets.
2. Choose Dataset: Select an available dataset for benchmarking.
3. Model Selection: Pick the deepfake detection models you want to benchmark your dataset against.
4. Start Benchmarking: Click the "Benchmark Dataset" button to initiate the benchmarking process.

The benchmarking results are displayed in detailed bar charts for DeepFake probability and inference time, and a downloadable CSV of the detection results is provided.

Adding Your Own Custom Dataset

1. Dataset Structure: Ensure your dataset follows this folder structure:

   datasets/
   └── image/ (or video/)
       └── your_dataset_name/
           ├── real/
           │   ├── image1.jpg
           │   └── ...
           └── fake/
               ├── image1.jpg
               └── ...
   

2. Config File: Create a .config file within your dataset folder to provide metadata about your dataset. Here’s an example of a .config file:

   [Dataset]
   name = Your Dataset Name
   description = A brief description of your dataset.
   source = URL or source information.

3. Upload Dataset: Place your dataset in the appropriate folder (datasets/image/ or datasets/video/).

4. Run Benchmark: Follow the steps in the benchmarking section to benchmark your custom dataset.

Future Work

DeepSafe acts as a platform where newer models can be incorporated into the app.

Contributing

Any kind of enhancement or contribution is welcomed. You can contribute your comments, questions, resources, and apps as issues or pull requests to the source code.

Acknowledgments

Methods	Repositories	Release Date
MesoNet	https://github.com/DariusAf/MesoNet	2018.09
FWA	https://github.com/danmohaha/CVPRW2019_Face_Artifacts	2018.11
VA	https://github.com/FalkoMatern/Exploiting-Visual-Artifacts	2019.01
Xception	https://github.com/ondyari/FaceForensics	2019.01
ClassNSeg	https://github.com/nii-yamagishilab/ClassNSeg	2019.06
Capsule	https://github.com/nii-yamagishilab/Capsule-Forensics-v2	2019.1
CNNDetection	https://github.com/peterwang512/CNNDetection	2019.12
DSP-FWA	https://github.com/danmohaha/DSP-FWA	2019.11
Upconv	https://github.com/cc-hpc-itwm/UpConv	2020.03
WM	https://github.com/cuihaoleo/kaggle-dfdc	2020.07
Selim	https://github.com/selimsef/dfdc_deepfake_challenge	2020.07
Photoshop FAL	https://peterwang512.github.io/FALdetector/	2019
FaceForensics	https://github.com/ondyari/FaceForensics	2018.03
CViT	https://github.com/erprogs/CViT	2021
Boken	https://github.com/beibuwandeluori/DeeperForensicsChallengeSolution	2020
GANimageDetection	GANimageDetection	License

License

This project is licensed under a dual license:

1. Open Source License (MIT License): For personal and non-commercial use.
2. Commercial License: For any commercial use, please contact Siddharth to obtain a commercial license.

Contact Information

For questions, please contact me at siddharth123sk[@]gmail.com

Made with ❤️ by Siddharth