README.md

Step 1: Generating the Perturbation

We first need to generate an image-agnostic Universal Adversarial Perturbation (UAP).

Quick Start

You may use anaconda or miniconda.

# Clone the repo
$ git clone https://github.com/wuhanstudio/adversarial-camera/
$ cd adversarial-camera/detection

# For CPU
$ conda env create -f environment.yml
$ conda activate adversarial-camera

# For GPU
$ conda env create -f environment_gpu.yml
$ conda activate adversarial-gpu-camera

$ python detect.py --model model/yolov3-tiny.h5 --class_name coco_classes.txt

The web page will be available at: http://localhost:9090/

That's it! The perturbation is saved as noise.npy file in the detection/noise/ folder.

White-box Adversarial Toolbox

Alternatively, you can generate the UAP using the WHite-box Adversarial Toolbox (WHAT).

Step 2: Deploying the Perturbation

You can use a raspberry pi 4 to deploy the perturbation (documentation).