Introduction

Pytorch implementation of "Beyond the Pixel World: A Novel Acoustic-based Face Anti-Spoofing System for Smartphones". (IEEE TIFS 2022)

Illustration of (a). the traditional RGB-based FAS system and (b). the proposed acoustic-based system.

Framework

Overview of the proposed two-branch Echo-FAS pipeline. The input signal is firstly fed forward to the Signal preprocessing module to extract nine signal segments. In the first branch, we apply Fast Fourier Transform (FFT) to obtain nine frequency signals, which are regarded as nine tokens and will be sent to the Transformer to obtain the global frequency feature. In the second branch, we further employ Short Time Fourier Transform (STFT) to obtain the corresponding spectrogram to the processed signal. The CNN is used to extract the local frequency feature of the input signals. The Cross Attention mechanisms model the correlation of the extracted global frequency feature $f_{1}$ and local frequency feature $f_{2}$. Subsequently, the attended features will be finally combined to determine whether the input query is a live person or a spoofer.

App

Coming soon.

Demo video

https://youtu.be/OoE5qFgwA54

Dataset

If you would like to access the Echo-Spoof dataset, please fill out this Google Form. The download link will be sent to you once the form is accepted.

Usage

1. Download Echo-Spoof database.
2. implement "python train_fusion.py" for reproducing the results. (change the training, validation, and test csv paths for different phones.)
3. implement "python test_fusion.py" for inference and evaluation. (change the test csv path and model path accordingly.)

FAS detection results for different smartphones

Phone	AUC(%)	ACC(%)	HTER(%)	EER(%)
Note	98.12	92.93	7.07	6.09
S9	99.86	97.57	2.43	1.24
S21	97.37	92.69	7.32	8.12
Xiaomi	99.81	97.51	2.45	2.14
AVG	98.79	95.18	4.83	4.34

Prerequisites:

• torch: 1.5.0
• torchvision: 0.6.0
• numpy: 1.23.0
• pandas: 1.0.5

Citation

@article{kong2022beyond,
  title={Beyond the Pixel World: A Novel Acoustic-based Face Anti-Spoofing System for Smartphones},
  author={Kong, Chenqi and Zheng, Kexin and Wang, Shiqi and Rocha, Anderson and Li, Haoliang},
  journal={IEEE Transactions on Information Forensics and Security},
  year={2022},
  publisher={IEEE}
}