CamPro: Camera-based Anti-Facial Recognition

Accepted by Network and Distributed System Security (NDSS) Symposium 2024

Requirements

Hardware

• more than 64 GB disk space
• at least one GPU that supports CUDA (Recommended: NVIDIA RTX 3090)

Software

• a recent Linux OS (e.g., Ubuntu 18.04/20.04/22.04)
• Anaconda
• CUDA driver whose version should be newer than 11.3

Semi-Auto Setup (Recommended)

• First, download all the three folders from the Google Drive and place them under this directory
• Second, run the script start.sh in your terminal. It will process the datasets and create the python virtual environment named CamPro.

Manually Setup

Python Environment Setup

We assume that you have installed the Python 3.9, and the CUDA Version should be higher than 11.3 (You can check this via nvidia-smi).

• If you have installed Anaconda, you could create a new environment via the command conda create --yes --name CamPro python=3.9.
• Then, you can activate the environment via the command conda activate CamPro.
• Finally, you should install the required packages via pip install -r requirements.txt.

Data Preparation

Google Drive: https://drive.google.com/drive/folders/1fvXBKqukA2BnGQU76QLtsRShuA5eiLY7?usp=sharing

• Datasets
  • CelebA: You should download the CelebA.zip from the Google Drive and unzip it under the datasets folder.
  • LFW: You should download the LFW.zip from the Google Drive and unzip it under the datasets folder.
  • COCO:
    • You should download the COCO.zip from the Google Drive and unzip it under the datasets folder. Then, make two empty folders, i.e., images and val2017, under the path datasets/COCO.
    • You should download COCO Detection 2017 images from the official website:
      • http://images.cocodataset.org/zips/train2017.zip
      • http://images.cocodataset.org/zips/val2017.zip
    • Extract all the JPEG images in train2017.zip to datasets/COCO/images.
    • Extract all the JPEG images in val2017.zip to datasets/COCO/val2017 and datasets/COCO/images.
• Models
  • checkpoints: You should download the checkpoints folder from the Google Drive.
  • weights: You should download the weights folder from the Google Drive.

Directories

After finishing data preparation, the CamPro directory should be organized like this below.

CamPro
├── checkpoints         # trained weights of various models
├── datasets            # open image datasets
│   ├── CelebA          # 112x112 cropped face dataset
│   ├── COCO            # person detection dataset
│   │   ├── images
│   │   ├── labels
│   │   ├── val2017
│   │   ├── val2017_mask
│   │   ├── test.txt
│   │   └── train.txt
│   └── LFW             # 112x96 cropped face dataset
├── results             # !!! collected results for artifact evaluation !!!
├── src                 # source codes
│   ├── baseline        # baseline methods of privacy protection
│   ├── evaluation      # !!! main python scripts for artifact evaluation !!!
│   ├── isp             # functions/classes related to ISP
│   ├── misc            # some helper functions
│   ├── privacy         # functions/classes related to privacy evaluation
│   ├── unet            # functions/classes related to image enhancer
│   ├── utility         # functions/classes related to utility evaluation
│   └── yolov5          # open-sourced object detection code repository
└── weights             # pre-trained weights of various models

Artifact Evaluation

In the following, we briefly introduce the execution steps. The details can be found in the AE appendix.

Anti-Facial Recognition Experiment (E1)

• conda activate CamPro
• cd src/
• python evaluation/exp1.py
• Results are saved to results/1.csv.

Ablation Study Experiment (E2)

• conda activate CamPro
• cd src/
• python evaluation/exp2.py
• Results are saved to results/2.csv.

Vision Application Performance Experiment (E3)

• conda activate CamPro
• cd src/
• python evaluation/exp3.py
• Results are saved to results/3.csv.

Image Quality Assessment Experiment (E4)

• conda activate CamPro
• cd src/
• python evaluation/exp4.py
• Results are saved to results/4.csv.

White-box Adaptive Attack Experiment (E5)

• conda activate CamPro
• cd src/
• python evaluation/exp5.py
• Results are printed in the console.