Dynamically Mitigating Data Discrepancy with Balanced Focal Loss for Replay Attack Detection

This repo contains the implementation of our work D3M towards building a more robust replay attack detection system. We propose an informative and complementary feature representation and leverage a more effective training objective. Experimental results in terms of min-tDCF and EER, as well as more detailed analysis will be reported in this repo.

Source code and other details for replay attack detection, tested on ASVspoof2019 PA and Real-PA dataset.

We are continuously adding comments and refining the repository. If you have some questions, feel free to open an issue:)

[x] source code of proposed methods
[x] attack samples for analysis
[x] model scores of seperate groups
[x] High-resolution images (in the near future)

Environment

apex 0.1 (for mixed precision training)
PyTorch 1.1.0 (DL framework)
sacred 0.7.5 (record experimental details)
Python 3.6+

To install most dependencies automatically:

pip install -r requirements.txt

Train the model

python main.py with 'epoch=50' 'lr=0.001'  'load_model=False' 'load_file=results/Model-epoch-25.pth' 'test_first=False' 'num_workers=1' 'eval_mode=False'

Test the model

python main.py with 'epoch=50' 'lr=0.001'  'load_model=False' 'load_file=results/models/best-eer-ep36-0.786008.pt' 'test_first=False' 'num_workers=1' 'eval_mode=True' 'server=0' 'train_batch=32' 'GRL_LAMBDA=0.001' 'evalProtocolFile=/data/to/anti-spoofing/ASVspoof2019/ASVspoof2019_PA_real/ASVspoof2019_PA_cm_protocols/ASVspoof2019.PA.real.cm.eval.trl.txt' 'eval_dir=/data/to/ASVspoof2019_PA_real/GDgram_magnitude_1024_400_240'

Use scared for experiment management

We use scared to manage our experiments, and you can create a file named myexp.py with your own configurations. For instance,

from sacred import Experiment
from sacred.observers import MongoObserver
from sacred.utils import apply_backspaces_and_linefeeds

ex = Experiment("ASVSPOOF2019")
ex.observers.append(MongoObserver.create(
    url='mongodb://exp:user@yourip:port/sacred?authMechanism=SCRAM-SHA-1',
    db_name='sacred'))
ex.captured_out_filter = apply_backspaces_and_linefeeds

Citation

If you find this work helpful, please cite it in your publications.

@inproceedings{dou2021dynamically,
author={Yongqiang Dou and Haocheng Yang and Maolin Yang and Yanyan Xu and Dengfeng Ke},
booktitle={The 25th International Conference on Pattern Recognition (ICPR)}, 
title={Dynamically Mitigating Data Discrepancy with Balanced Focal Loss for Replay Attack Detection}, 
year={2021},
volume={},
number={},
pages={4115-4122},
doi={10.1109/ICPR48806.2021.9412749}}

asvspoof / D3M

readme