LiVSec - MMSys'23

The LiVSec project aims to investigate the security implications of live 3D video surveillance. We identified the security vulnerabilities caused by the 3D sensitive objects (e.g., human face) in the surveillance video that lead to spoofing attacks. To address the security issue, we developed a novel perturbation generation method, namely LiVSec, which employs a real-time generative neural network model to inject small perturbations to the 3D video frames to blur and protect the sensitive objects.

The details of this project can be found in our accepted MMSys'23 paper:

Zhongze Tang, Huy Phan, Xianglong Feng, Bo Yuan, Yao Liu, and Sheng Wei. 2023. Security-Preserving Live 3D Video Surveillance. In Proceedings of the 14th ACM Multimedia Systems Conference (MMSys ’23), June 7–10, 2023, Vancouver, BC, Canada. https://doi.org/10.1145/3587819.3590975

The paper can be found under the paper folder.

This repo contains both code and instructions for the following three components:

1. Reproduce the experimental results reported in the paper.
2. Train your own 3D face authentication model and the real-time perturbation generator to prevent the face models in the surveillance video from being exploited to spoof the face authentication.
3. Set up the end-to-end security-preserving live 3D video surveillance system integrating the perturbation generator.

Repository Hierarchy

├── FaceAuthentication              // Face authentication system
...
├── ProtectedFaceGeneration         // The core of LiVSec, the generative model that adds protection to 3D video surveillance
│   ├── modules                     // Where you should put two pre-trained models
...
├── System                          // Security-preserving live 3D video surveillance system
│   ├── modules                     // Where you should put two pre-trained models
│   ├── John_3dts.mp4               // The source of Dataset #2 w/ timestamp.
...
├── paper
├── LICENSE
├── LiVSec_MMSys23_ReproducibilityAppendix.pdf
└── README.md

1. Environmental Setup

Hardware requirements

A workstation with an Nvidia GPU is required to train/infer the models, and run the system.

In our project, an Nvidia RTX A6000 GPU is used to evaluate the system.

Software requirements

The project is developed and tested in the following environment:

Ubuntu 20.04 LTS
CUDA 11.8
Python 3.7
PyTorch 1.7.1+cu110
PyTorch-Lightning 1.2.3

Please set up the development environment by following the instructions below:

1. Update Ubuntu first.

   sudo apt update
   sudo apt upgrade

2. Install CUDA (a version >=11.0 && <12.0 is okay) following the official instructions.

3. (Optional) Install Anaconda to setup a virtual environment. This will help protect your dev environment from being a mess. Set up the Conda virtual environment w/ Python 3.7. We highly recommend using PyCharm GUI to do so directly. If you prefer the command line, try the following commands to create a virtual environment called LiVSec_MMSys23.

   conda create -n LiVSec_MMSys23 python=3.7
   conda activate LiVSec_MMSys23

4. Install the required Python libraries (in the virtual environment if you have one).

   pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
   pip install matplotlib pytorch-lightning==1.2.3 wandb 
   pip install lpips opencv-python==4.3.0.38

5. [For System] Install ffmpeg.

   sudo apt install ffmpeg

   Please note that in this project, we assume the path to the binary file of ffmpeg is /usr/bin/ffmpeg. You can find its path in your system by typing which ffmpeg in the terminal. If it is different from this path, please modify Line 30 of main.py under System folder to the ffmpeg path in your system.

6. [For System] Install v4l2loopback following the official instructions.

   Make sure to enable it before running the LiVSec system.

   sudo modprobe v4l2loopback

   v4l2loopback by default will create a virtual camera called /dev/video0, run the following code, and if it is created successfully, the output will be video0. Refer to its official documentation for more details.

   ls /dev | grep video0

7. Download the models from https://drive.google.com/drive/folders/17WVDVuHnQpau84fJzXwZFBzaGKyjUEE5?usp=share_link.

   Put two pre-trained models under ./ProtectedFaceGeneration/modules and ./System/modules.

2. Reproduce the Results of LiVSec

Prepare datasets

Dataset #1 comes from https://vap.aau.dk/rgb-d-face-database/. Run preprocess_data.py under FaceAuthentication/utils to download and preprocess Dataset #1.

If you want to change where to put the dataset, change Line 126 of preprocess_data.py. The default is /data/faceid. If you change it, you should also change --data_path argument every time you call the models (either the face authentication model or the protected face generation model).

The 3D video in Dataset #2 is exported from Depthkit (See Appendix section of this doc for more details). We provide the exported video we use in this project, which is System/John_3dts.mp4. No additional actions are needed for Dataset #2 since they are all included.

Results for Dataset #1

Run result_collect.py under ProtectedFaceGeneration folder. The output will show all the results.

Results for Dataset #2

Run main.py under System folder. The output will show all the results in REUSE-1 mode.

Change Line 28 of main.py to the designated reuse frequency, e.g., 5 or 10, which was used in the paper, to produce corresponding results.

Please note that to see the results, the MODE in Line 27 has to be set as RESULT_COLLECT_MODE.

3. [Optional] Train Your Own Models

See the 7th step of Environmental Setup to download the pretrained models. You can also train your own face authentication and protected face generation models.

You should follow Prepare datasets section above to prepare the datasets first.

Train the face authentication model

Run train.py under FaceAuthentication folder. If you want to change the training hyper-parameters, check Line 41-59 of the train.py.

Train the protected face generation model

You should have a pretrained face authentication model, and put it under ProtectedFaceGeneration/modules. Change Line 21 of ProtectedFaceGeneration/main_module.py to your own face authentication model path.

Run train.py under ProtectedFaceGeneration folder. If you want to change the training hyper-parameters, check Line 43-59 of the train.py.

4. [Optional] The End-to-end Security-preserving Live 3D Video Surveillance System

The protected frames will be sent to the virtual camera /dev/video0. Any end-to-end live streaming protocols can be adopted to build the live streaming system. In this project, we use DashCast to generate a live DASH stream from it at the server end, and use MP4Client to watch the playback at the client end.

Server-end configuration

The code under System folder runs on the server-end, so you should do the following at the server end.

1. Install DashCast (included in the GPAC 0.8.1). Find more versions about GPAC 0.8.1, see this page.

   wget https://download.tsi.telecom-paristech.fr/gpac/legacy_builds/linux64/gpac/gpac_0.8.1-latest-legacy_amd64.deb
   sudo apt install ./gpac_0.8.1-latest-legacy_amd64.deb

2. Install Node.js. We recommend using tools like nvs to easily install and manage Node.js. See official documentation of nvs for more information.

   # install nvs
   export NVS_HOME="$HOME/.nvs"
   git clone https://github.com/jasongin/nvs "$NVS_HOME"
   . "$NVS_HOME/nvs.sh" install
   
   # add LTS version of the node
   nvs add lts

3. Download the DASH Low Latency Web Server.

   git clone https://github.com/gpac/node-gpac-dash
   cd node-gpac-dash && mkdir livsec

   We assume the path of the web server is <YOUR_PATH>/node-gpac-dash. Change Line 34 of Server/main.py to <YOUR_PATH>/node-gpac-dash/livsec.

4. Start DASH streaming. We provide two types of DASH, the default one and the low-latency one. Start the corresponding DASH server.

   cd <YOUR_PATH>/node-gpac-dash
   nvs use lts  # enable node
   
   # for default DASH
   node gpac-dash.js
   
   # for low latency DASH
   node gpac-dash.js -segment-marker eods -chunk-media-segments

   Change the DASH_MODE parameter in Line 33 of System/main.py to switch the DASH streaming mode. You also need to change Line 27 of it to STREAMING_MODE to disable the result collection. Finally, run main.py, and the DASH profiles/segments generation will start automatically. You can check the generated DASH files in <YOUR_PATH>/node-gpac-dash/livsec/output.

Client-end configuration

1. Install MP4Client by installing GPAC 0.8.1 following the aforementioned steps.
2. Watch the playback.

   # default DASH
   MP4Client http://localhost:8000/livsec/output/manifest.mpd
   
   # low latency DASH
   MP4Client http://127.0.0.1:8000/livsec/output/dashcast.mpd -opt Network:BufferLength=200 -opt DASH:LowLatency=chunk -opt DASH:UseServerUTC=no

Known issues

For both DASH streaming modes, the streaming is not smooth. And in the low-latency mode, there is a distortion at the client end. We will keep working on this and provide updates in this GitHub repo.

5. Appendix

How to get Dataset #2

1. Get a Depthkit account at https://www.depthkit.tv/signup.
2. Log in and download Depthkit (as of the paper submission, the version is 0.6.1), and the Sample Project for Depthkit Core (which contains a pre-shot volumetric video) at https://www.depthkit.tv/downloads.
3. Install Depthkit and unzip the Sample Project for Depthkit Core. Use Depthkit to open the sample project.
4. Click the Editor button (it looks like a film) in the upper-left corner of the window to begin editing the sample project.
5. Select the ISOLATE tab, and change Depth Range to the range of 1.27m to 2.37m. This step filters foreground and background.
6. Change all numbers for Horizontal Crop and Vertical Crop to 0px.
7. Use the two selectors (look like label icons) in the lower-right of the window to select the video clip with the length you desire.
8. Go to the EXPORT tab, and set the Output Width and Output Height to 2048.
9. Click the EXPORT button, and you will get an RGBD video similar to Figure 5 in the paper.

Please note that, while conducting the experiments for this paper, we used an older version of Depthkit (around late 2019 or early 2020) to edit and export the RGBD video. Now the newest version of Depthkit cannot generate the same RGBD video clip. For example, in Step 8, when using the current version of Depthkit, the Output Width and Output Height can only be set to 1920 and 2160, respectively.

6. Cite Our Work

Zhongze Tang, Huy Phan, Xianglong Feng, Bo Yuan, Yao Liu, and Sheng Wei. 2023. Security-Preserving Live 3D Video Surveillance. In Proceedings of the 14th ACM Multimedia Systems Conference (MMSys ’23), June 2023. https://doi.org/10.1145/3587819.3590975

7. Contact

If you have any questions or any idea to discuss, you can email Zhongze Tang ([email protected]) directly.

8. License

MIT