HaTFAS

Official repository for the article "Efficient Face Anti-Spoofing via Head-aware Transformer based Knowledge Distillation with 5 MB model parameters", which has been accepted at Applied Soft Computing and can be found here.

---

🔔Introduction

Inspired by the powerful global modelling ability of self-attention and the model compression ability of knowledge distillation, a simple yet effective knowledge distillation approach is proposed for Face Anti-Spoofing(FAS) under transformer framework. Our primary idea is to leverage the rich knowledge of a teacher network pre-trained on large-scale face data to guide the learning of a lightweight student network. The main contributions of our method are threefold: (1) Feature- and logits-level distillation are combined to transfer the rich knowledge of teacher to student. (2) A head-aware strategy is proposed to deal with the dimension mismatching issue of middle encoder layers between teacher and student networks, in which a novel attention head correlation matrix is introduced. (3) Our method can bridge the performance gap between teacher and student, and the resulting student network is extremely lightweight with only 5 MB parameters.

🔨 Installation

• Install CUDA 12.0 with cuDNN 8 following the official installation guide of CUDA and cuDNN.

• Setup conda environment:

# Create environment
conda create -n HaTFAS python=3.8.20
conda activate HaTFAS

# Install requirements
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118

# Clone HaTFAS
git clone https://github.com/Maricle-zhangjun/HaTFAS.git
cd HaTFAS

# Install other requirements
pip install -r requirements.txt

🚀 Getting started

Data preparation

Since the datasets used in our research are face videos, each video is first split into frames. Then, face region of each video frame is detected via MTCNN and resized the size of 224×224×3. The folder of pro-processed dataset as:

dataset_path/train(val or test)/true/video_dir/face_image_1.jpg
dataset_path/train(val or test)/attack/video_dir/face_image_1.jpg

Pretraining teacher networks

For each face dataset, teacher network is pre-trained based on DeiT and then used to supervise the learning of student, the backbone differences between teacher and student as:

Model	Embedding dimension	Heads	Layers
Teacher (ViT-Base)	768	12	12
Student (ViT-Tiny)	192	3	12

Distilling student networks

To distill student based on ViT-Tiny with HaTFAS, run:

python main.py --batch-size 128 --epochs 300 --distributed --output_dir <model_save_path>\
    --data-path <face_dataset_path> --data-set FACE --teacher-path <pretrained_teacher_path>\
    --model deit_tiny_patch16_224 --teacher-model deit_base_patch16_224 --distillation-type soft\
    --distillation-alpha 0.7 --distillation-beta 1 --distillation-tau 3 --s-id 4 6 8 --t-id 4 6 8

Evaluating student networks

To eval the distilled student with HaTFAS, run:

python main.py --batch-size 128 --eval --data-path <face_dataset_path> --data-set FACE_TEST\
    --model deit_tiny_patch16_224 --resume <distilled_student_path>

Main results

Comparison results on CASIA-FASD and Replay-Attack datasets

Comparison results with state-of-the-art methods on OULU-NPU dataset

Comparison results with knowledge distillation methods on OULU-NPU dataset.

✏️ Citation

If you find this work useful for your research, please feel free to leave a star⭐️ and cite our paper:

@article{ZHANG2024112237,
title = {Efficient face anti-spoofing via head-aware transformer based knowledge distillation with 5 MB model parameters},
journal = {Applied Soft Computing},
volume = {166},
pages = {112237},
year = {2024},
issn = {1568-4946},
doi = {https://doi.org/10.1016/j.asoc.2024.112237},
author = {Jun Zhang and Yunfei Zhang and Feixue Shao and Xuetao Ma and Shu Feng and Yongfei Wu and Daoxiang Zhou}
}