______                   ___          __  _    _____                   _____
   / ____/___ _________     /   |  ____  / /_(_)  / ___/____  ____  ____  / __(_)___  ____ _
  / /_  / __ `/ ___/ _ \   / /| | / __ \/ __/ /   \__ \/ __ \/ __ \/ __ \/ /_/ / __ \/ __ `/
 / __/ / /_/ / /__/  __/  / ___ |/ / / / /_/ /   ___/ / /_/ / /_/ / /_/ / __/ / / / / /_/ /
/_/    \__,_/\___/\___/  /_/  |_/_/ /_/\__/_/   /____/ .___/\____/\____/_/ /_/_/ /_/\__, /
                                                    /_/                            /____/

Problem Description

This project aims to evaluate the performance of lightweight face models for facial anti-spoofing (FAS), comparing their accuracy and computational complexity with state-of-the-art deep models.

Input: Image with face

Output: Fake or Real

Dataset

LCCD FASD Dataset. Link: https://www.kaggle.com/datasets/faber24/lcc-fasd

Methods

• ResNeXT50

• MobileNetV3

• FeatherNet

Results

We evaluate on LCCD FASD Development after preprocessing. Link: CV_dataset

Model	APCER	NPCER	ACER
ResNeXT50	0.1308	0.3037	0.2127
MobileNetV3	0.1727	0.2111	0.1917
FeatherNet	0.1994	0.1284	0.1639

Installation

1. Clone Project

git clone https://github.com/dtruong46me/face-anti-spoofing.git
cd face-anti-spoofing

2. Install requirements.txt

bash setup.sh

3. Training First, you should download CV_dataset and put it into the folder /face-anti-spoofing/cv-dataset and run the scipt:

python run_training.py \
--train_path "cv-dataset/final_data/train" \
--test_path "cv-dataset/final_data/valid" \
--batch_size 128 \
--modelname "seresnext50"\
--wandb_token "<your_wandb_token>" \
--wandb_runname "<your_wandb_run_name>" \
--num_classes 2 \
--max_epochs 40

or your can use bash train_all.sh to train, evaluate, predict all pretrained models.

You can follow scipts in the notebook: https://www.kaggle.com/code/dtruon46/master-face-anti-spoofing or the file: /face-anti-spoofing/code train.ipynb

4. Demo

You can download weights of models (.ckpt file) and put it into the /face-anti-spoofing/FAS_detector/model/your_model.ckpt. Put your test image in the /face-anti-spoofing/assets/samples/your_images.jpg

• ResNeXT50: Weights (106Mb)
• MobileNetV3: Weights (28MB)
• FeatherNet: Weights (4.3MB)

Then run the scripts

python predict_sample.py \
--model_checkpoint "your_model.ckpt" \
--image "your_test_image.jpg"\
--modelname "seresnext50"

Contributions

• Supervisor: Prof. Dang Tuan Linh

• Group Members

No.	Name	Student ID	Email
1	Vu Tuan Minh (Leader)	20210597	minh.vt210597@sis.hust.edu.vn
2	Nguyen Tien Doanh	20214881	doanh.nt214881@sis.hust.edu.vn
3	Nguyen Tung Luong	20214913	luong.nt214913@sis.hust.edu.vn
4	Phan Dinh Truong	20214937	truong.pd214937@sis.hust.edu.vn