VVSec is a novel volumetric video security mechanism, which makes benign use of adversarial perturbations to obfuscate the security and privacy sensitive 3D face models. Such obfuscation ensures that the 3D models cannot be exploited to bypass deep learning-based face authentications. Meanwhile, the injected perturbations are not perceivable by the end-users, maintaining the original quality of experience in volumetric video streaming.
To find more details, please read our paper:
Zhongze Tang, Xianglong Feng, Yi Xie, Huy Phan, Tian Guo, Bo Yuan, and Sheng Wei. 2020. VVSec: Securing Volumetric Video Streaming via Benign Use of Adversarial Perturbation. In Proceedings of the 28th ACM International Conference on Multimedia (MM '20). Association for Computing Machinery, New York, NY, USA, 3614–3623. DOI: https://doi.org/10.1145/3394171.3413639
This repository contains the source code and the dataset to test our VVSec, which adds perturbation to a human face in a volumetric video so that the face and its corresponding depth can no longer be used to bypass a deep learning-based face authentication system.
Dataset #1 has already been included in this repo, under ./dataset/ds1/.
Dataset #2 can be found at https://vap.aau.dk/rgb-d-face-database/.
You can check Datasets section for details.
Download the model of the face authentication system
from https://drive.google.com/file/d/16v77oJZafuFbw8_vuoQ2jwmx0_N_Q6y5/view?usp=sharing,
and put it under ./model/.
You can check Face Authentication System section for details.
We recommend using Anaconda to setup the environment.
Install the following libraries:
- Python 3.7
- keras 2.2.4
- TensorFlow 1.14.0
- Pillow 6.2.1
- numpy 1.17.4
- matplotlib 3.3.1
Simply run main.py to see the results!
If you want to change the inputs, just simply modify the path of the reference and user inputs files.
These two figures match Case 1 and 9 of Table 1 in our paper, respectively.
.
├── adv_attack.py // Adversarial attack algorithm
├── dataset
│ ├── ds1 // Dataset 1, which comes from a Volumetric Video
│ │ ├── 0.jpg_c.bmp // RGB image used for VVSec
│ │ ├── 0.jpg_d.bmp // Visualized depth data
│ │ ├── 0.jpg_d.dat // Depth data used for VVSec
│ │ ├── ...
│ └── ds2 // Dataset 2, the RGB-D dataset
│ ├── faceid_train
│ │ ├── (2012-05-16)(151751)
│ │ │ ├── 001_1_c.bmp // RGB image used for VVSec
│ │ │ ├── 001_1_d.dat // Depth data used for VVSec
│ │ │ ├── 001_2_c.bmp
│ │ │ ├── 001_2_d.dat
│ │ │ ...
│ │ ...
│ ├── faceid_val
│ │ ├── (2012-05-18)(152717)
│ │ │ ├── 001_1_c.bmp
│ │ │ ├── 001_1_d.dat
│ │ │ ├── 001_2_c.bmp
│ │ │ ├── 001_2_d.dat
│ │ │ ...
│ │ ...
│ └── PUT_DATASET2_HERE
├── main.py // Main program, quickly try VVSec
├── model
│ └── model_top.model // Model for the face authentication used in VVSec, you have to download it
├── pic
│ ├── demo_result_1.png
│ └── demo_result_2.png
├── README.md // This file
├── utils.py // Some helper functions
└── .gitignore
For both datasets, the RGB images are stored as a 32-bit bitmap (*_c.bmp) with resolution 1280 × 960px.
The depth images are stored as plaintext files (*_d.dat), where a depth value represents the corresponding depth
pixel, with resolution 640×480px. The numbers in the depth data files are distance in millimetres from the sensor.
Other values are constants specified by the Kinect:
- -1 - Undefined distance
- 0 - Too near
- 4095 - Too far
Dataset #1 contains 11 RGB-D images of one user extracted from the volumetric video demo in the Depthkit.
To download the volumetric video, you need to log into DepthKit to visit https://www.depthkit.tv/downloads, then
download the Sample Project.
We also provide visualized depth information of all the data in Dataset #1, which ends in *_d.bmp.
It comes from the paper An RGB-D Database Using Microsoft's Kinect for Windows for Face Detection, and its original download webpage is https://vap.aau.dk/rgb-d-face-database/.
We adopt an open-source 3D image-based face authentication system to validate the attack and defense mechanisms presented in this work.
We provide the file of the DNN model in this Siamese Network, which can be downloaded from
https://drive.google.com/file/d/16v77oJZafuFbw8_vuoQ2jwmx0_N_Q6y5/view?usp=sharing, and should be put under ./model/.
To get the complete model of
this face authentication system, call get_model() function in utils.py.
@inproceedings{tang20vvsec,
title = {VVSec: Securing Volumetric Video Streaming via Benign Use of Adversarial Perturbation},
author = {Tang, Zhongze and Feng, Xianglong and Xie, Yi and Phan, Huy and Guo, Tian and Yuan, Bo and Wei, Sheng},
booktitle = {ACM International Conference on Multimedia (MM)},
pages = {3614--3623},
year = {2020}
}If you have any questions or ideas to discuss, please contact Zhongze Tang ([email protected]). Thank you!
MIT