Guarding Against Deepfake Audio with One-Class Softmax

The codebase in this project is adapted from AIR-ASVspoof and deepfake-whisper-features.

The project aims to explore the generalization capability of anti-spoofing systems utilizing the One-Class Softmax loss function (Zhang et al., 2021) on real-world samples. Please refer to the report and poster for more details.

• This repo: Various experiments of ResNet18.
• Whisper-ocsoftmax: Using this repo to run experiment of MFCC+Whisper+MesoNet-OCSoftmax.
• Whisper-softmax: Using this repo to run experiment of MFCC+Whisper+MesoNet-Softmax.

Requirements

python==3.10 pytorch==2.3

Data Preparation

The LFCC features are extracted with the MATLAB implementation provided by the ASVspoof 2019 organizers.

• ASVSpoof2019: first run the process_LA_data.m with MATLAB, and then run python3 reload_data.py.
• In-the-wild: first run the process_in_the_wild_data.m with MATLAB, and then run python3 reload_data.py. Make sure to change the directory path to the path on your machine.

Run the training code

Before running the train.py, change the path_to_features, path_to_protocol according to the files' location on your machine.

E.g., LFCC-ResNet18 on ASVspoof_2019:

CUDA_VISIBLE_DEVICES=[idx] python train.py --dataset_version ASVspoof_2019 --
add_loss [softmax/ocsoftmax] --frontend lfcc -o ./models/[softmax/ocsoftmax] --batch_size=32 --num_epochs 20

Recommended configuration for ASVspoof_2021:

CUDA_VISIBLE_DEVICES=[idx] python train.py --add_loss ocsoftmax -o ./models_asv21_mfcc_eer_1004/ocsoftmax/ --batch_size=8 --train_amount 100000 --valid_amount 25000

Run the test code with trained model

Change the model_dir to the location of the model you would like to test with.

python test.py -m ./models_asv19_lfcc_20ep/ocsoftmax -l ocsoftmax -t in_the_wild