ARCFACE Face Recognition - Attack & Defense System

This project implements a comprehensive face recognition attack and defense system using the ARCFACE model.

Features

Supported Attack Methods

1. BIM (Basic Iterative Method)
2. DIM (Diverse Input Method)
3. TIM (Translation-Invariant Method)
4. PGD (Projected Gradient Descent)
5. CW (Carlini & Wagner)
6. DeepFool

Supported Defense Methods

1. HGD (High-level Guided Denoising)
2. TVM (Total Variation Minimization)
3. LivenessDetection (Face Spoofing Detection)
4. FeatureSpacePurification
5. EnsembleDefense

Supported Datasets

1. VGGFace2
2. CelebA
3. CASIA-WebFace
4. LFW (Labeled Faces in the Wild)
5. YaleB
6. MegaFace

Project Structure

nudt_arcface/
├── attacks/                # Attack methods implementation
│   ├── bim.py
│   ├── dim.py
│   ├── tim.py
│   ├── pgd.py
│   ├── cw.py
│   └── deepfool.py
├── defends/               # Defense methods implementation
│   ├── hgd.py
│   ├── tvm.py
│   ├── liveness_detection.py
│   ├── feature_space_purification.py
│   └── ensemble_defense.py
├── datasets/              # Dataset loaders
│   ├── vggface2.py
│   ├── celeba.py
│   ├── webface.py
│   ├── lfw.py
│   ├── yaleb.py
│   └── megaface.py
├── face_recognizer/       # Model implementation
│   ├── arcface_model.py
│   └── main.py
├── utils/                 # Utility functions
│   ├── sse.py            # SSE output formatting
│   └── yaml_rw.py        # YAML config handling
├── main.py               # Entry point
├── Dockerfile            # Docker image configuration
├── requirements.txt      # Python dependencies
├── docker_run_scripts.sh # Docker run examples
└── ENV_VARIABLES.md      # Environment variables documentation

Installation

Local Installation

# Install PyTorch
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu121

# Install dependencies
pip install -r requirements.txt

Docker Installation

# Build Docker image
docker build -t nudt_arcface:latest .

Usage

Running Locally

Attack Mode

python main.py \
  --input_path ./input \
  --output_path ./output \
  --process adv \
  --attack_method bim \
  --epsilon 0.031 \
  --step_size 0.008 \
  --max_iterations 10 \
  --device 0

Defense Mode

python main.py \
  --input_path ./input \
  --output_path ./output \
  --process defend \
  --defend_method hgd \
  --device 0

Running with Docker

BIM Attack Example

docker run --rm --gpus all \
  -v /path/to/input:/project/input:ro \
  -v /path/to/output:/project/output:rw \
  -e PROCESS=adv \
  -e ATTACK_METHOD=bim \
  -e EPSILON=0.031 \
  -e STEP_SIZE=0.008 \
  -e MAX_ITERATIONS=10 \
  -e DEVICE=0 \
  nudt_arcface:latest

HGD Defense Example

docker run --rm \
  -v /path/to/input:/project/input:ro \
  -v /path/to/output:/project/output:rw \
  -e PROCESS=defend \
  -e DEFEND_METHOD=hgd \
  nudt_arcface:latest

See docker_run_scripts.sh for more examples.

Input Directory Structure

input/
├── model/
│   └── model_weights.pth    # Pretrained model weights
└── data/
    └── dataset_name/         # Dataset folder
        ├── image1.jpg
        ├── image2.jpg
        └── ...

Output Format

SSE (Server-Sent Events) Format

All outputs follow the SSE format for real-time progress monitoring:

event: input_path_validated
data: {"status": "success", "message": "Input path is valid and complete.", "file_name": "/project/input"}

event: adv_samples_gen_validated
data: {"status": "success", "message": "adv sample is generated.", "file_name": "/project/output/adv_image1.jpg"}

Environment Variables

See ENV_VARIABLES.md for complete documentation of all supported environment variables.

Key Variables

Variable	Description	Default	Example
PROCESS	Operation mode	attack	adv, attack, defend, train
ATTACK_METHOD	Attack to use	bim	bim, dim, tim, pgd, cw, deepfool
DEFEND_METHOD	Defense to use	hgd	hgd, tvm, livenessdetection, etc.
EPSILON	Attack strength	0.031	0.0 - 1.0
MAX_ITERATIONS	Attack iterations	10	1 - 1000
DEVICE	GPU device ID	0	-1 (CPU) or 0+ (GPU)

Model Information

ARCFACE Architecture

• Backbone: ResNet-50
• Embedding Size: 512 dimensions
• Input Size: 112x112 pixels
• Output: Identity classification logits

Examples

Example 1: PGD Attack on LFW Dataset

docker run --rm --gpus all \
  -v /data/lfw:/project/input:ro \
  -v /data/output:/project/output:rw \
  -e PROCESS=adv \
  -e DATA=lfw \
  -e ATTACK_METHOD=pgd \
  -e EPSILON=0.062 \
  -e STEP_SIZE=0.016 \
  -e MAX_ITERATIONS=20 \
  -e DEVICE=0 \
  nudt_arcface:latest

Example 2: Ensemble Defense

docker run --rm \
  -v /data/adversarial:/project/input:ro \
  -v /data/defended:/project/output:rw \
  -e PROCESS=defend \
  -e DEFEND_METHOD=ensembledefense \
  nudt_arcface:latest

Example 3: C&W Attack with Custom Parameters

docker run --rm --gpus all \
  -v /data/celeba:/project/input:ro \
  -v /data/output:/project/output:rw \
  -e PROCESS=adv \
  -e DATA=celeba \
  -e ATTACK_METHOD=cw \
  -e MAX_ITERATIONS=100 \
  -e DEVICE=0 \
  nudt_arcface:latest