The code is heavily inspired by the AmI implementation and cleverhans implementation.
- Creating Secure REST API in Node.js (run a very simple HTTP server) to provide online services to have public-facing APIs.
- Using child_process module of Node.js to spawn subprocesses of command by batch (.bat).
- Switching conda virtual environment and programs by batch.
- Every GET request to the Node.js server will get corresponding result by a HTML page.
| API No. | API | Method | Functionality |
|---|---|---|---|
| 1 | /facenet?{attack}&{victim} |
GET | implement face verification between attacker's face and victim's face |
| 2 | /adv |
GET | implement fgsm attack between attacker's face and victim's face and then implement face verification |
| 3 | /detect |
GET | implement adversary detection between attacker's face and victim's face and give result |
There are four sub-folders under /demo which are: photo, origin, adv, detect.
| folder | Content | Changed by API |
|---|---|---|
| photo | original face images of attacker and victim | NaN |
| origin | facenet verification result plot | /facenet?{attack}&{victim} |
| adv | attacker's face with adversarial noise and facenet verification result plot | /adv |
| detect | adversary detection result plot | /detect |
-
Please download VGG-Face caffe model and unzip the model under
model/folder. -
Please download facenet pre-trained model 20180402-114759 and unzip the model under
model/folder.
As we use VGG-Face caffe model to detect adversary, so we need to compile caffe before detecting adversary. Please click here to compile.
| Environment | Purpose |
|---|---|
| Visual Studio 2015 + Anaconda env(python 3.5) + CUDA 8.0 + cuDNN 5 | Compile caffe for detecting adversary |
| Visual Studio 2015 + Anaconda env(python 3.6) + tensorflow-gpu 1.7.0 + CUDA 9.0 + cuDNN 7.0.5 | For runing face verfication (using gpu) and fgsm attack |
| Visual Studio 2015 + Anaconda env(python 3.5) + tensorflow-gpu 1.7.0 + CUDA 9.0 + cuDNN 7.0.5 | For runing detection part (using gpu) |
| Visual Studio 2015 + Anaconda env(python 3.5) + tensorflow 1.7.0 + CUDA 9.0 + cuDNN 7.0.5 | For runing detection part (using cpu) |
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\libnvvp
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\lib\x64
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\include
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\extras\CUPTI\lib64
C:\Program Files\NVIDIA Corporation\NVSMI
C:\Program Files\NVIDIA Corporation\NVIDIA NvDLISR
C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common
$nvcc -V # to see if GPU is configured successfully
$vidia-smi.exe -l 1 # update GPU information every second$conda create -n attack python=3.6
$conda activate attack
$pip install -r requirements.txt
$pip install cleverhans==3.0.1 tensorflow-gpu==1.7.0 scikit-learn opencv-python numpy==1.16.2 scipy==1.2.1 matplotlib==3.1.1 Pillow pytest-shutil
$python origin\origin.py Elon_Musk Jeff_Bezos
$python adv/fgsm.py$conda create -n detect python=3.5
$conda activate detect
$pip install scikit-image scipy==1.2.1 opencv-python scikit-learn matplotlib Pillow tensorflow-gpu==1.7.0 protobuf==3.4.0
$conda install protobuf
# copy caffe\python\caffe to conda_virtual_env\Lib\site-packages
$python detect/detect.py$pip freeze > requirements.txt
$conda env export > environment.yaml
# $conda activate xx && $pip install -r requirements.txt
# $conda env create -f environment.yamlFirstly, run Node.js server by the following command to start the http server:
$npm install
$node server.jsSecondly, open web browser to send your GET requests as follows:
http://127.0.0.1:3000/facenet?Elon_Musk&Jeff_Bezos
http://127.0.0.1:3000/adv
http://127.0.0.1:3000/detect
Contributions are always welcome! Feel free to dive in!
Please read the contribution guideline first, then open an issue open an issue or submit PRs.
This repository follows the Contributor Covenant Code of Conduct.
This project exists thanks to all the people who contribute.
MIT © Elaine Zhong