WaveFake: A Data Set to Facilitate Audio DeepFake Detection

logo

This is the code repository for our NeurIPS 2021 (Track on Datasets and Benchmarks) paper WaveFake.

Deep generative modeling has the potential to cause significant harm to society.
Recognizing this threat, a magnitude of research into detecting so-called “Deepfakes” has emerged.
This research most often focuses on the image domain, while studies exploring generated audio signals have – so far – been neglected.
In this paper, we aim to narrow this gap.
We present a novel data set, for which we collected ten sample sets from six different network architectures, spanning two languages.
We analyze the frequency statistics comprehensively, discovering subtle differences between the architectures, specifically among the higher frequencies.
Additionally, to facilitate further development of detection methods, we implemented three different classifiers adopted from the signal processing community to give practitioners a baseline to compare against.
In a first evaluation, we already discovered significant trade-offs between the different approaches.
Neural network-based approaches performed better on average, but more traditional models proved to be more robust.

Dataset & Pre-trained Models

You can find our dataset on zenodo and we also provide pre-trained models.

Setup

You can install all needed dependencies by running:

pip install -r requirements.txt

RawNet2 Model

For consistency, we use the RawNet2 model provided by the ASVSpoof 2021 challenge.
Please download the model specifications here and place it under dfadetect/models as raw_net2.py.

Statistics & Plots

To recreate the plots/statistics of the paper, use:

python statistics.py -h

usage: statistics.py [-h] [--amount AMOUNT] [--no-stats] [DATASETS ...]

positional arguments:
  DATASETS              Path to datasets. The first entry is assumed to be the referrence one. Specified as follows <path,name>

optional arguments:
  -h, --help            show this help message and exit
  --amount AMOUNT, -a AMOUNT
                        Amount of files to concider.
  --no-stats, -s        Do not compute stats, only plots.

Example

python statistics.py /path/to/reference/data,ReferenceDataName /path/to/generated/data,GeneratedDataName -a 10000

Training models

You can use the training script as follows:

python train_models.py -h

usage: train_models.py [-h] [--amount AMOUNT] [--clusters CLUSTERS] [--batch_size BATCH_SIZE] [--epochs EPOCHS] [--retraining RETRAINING] [--ckpt CKPT] [--use_em] [--raw_net] [--cuda] [--lfcc] [--debug] [--verbose] REAL FAKE

positional arguments:
  REAL                  Directory containing real data.
  FAKE                  Directory containing fake data.

optional arguments:
  -h, --help            show this help message and exit
  --amount AMOUNT, -a AMOUNT
                        Amount of files to load from each directory (default: None - all).
  --clusters CLUSTERS, -k CLUSTERS
                        The amount of clusters to learn (default: 128).
  --batch_size BATCH_SIZE, -b BATCH_SIZE
                        Batch size (default: 8).
  --epochs EPOCHS, -e EPOCHS
                        Epochs (default: 5).
  --retraining RETRAINING, -r RETRAINING
                        Retraining tries (default: 10).
  --ckpt CKPT           Checkpoint directory (default: trained_models).
  --use_em              Use EM version?
  --raw_net             Train raw net version?
  --cuda, -c            Use cuda?
  --lfcc, -l            Use LFCC instead of MFCC?
  --debug, -d           Only use minimal amount of files?
  --verbose, -v         Display debug information?

Example

To train all EM-GMMs use:

python train_models.py /data/LJSpeech-1.1/wavs /data/generated_audio -k 128 -v --use_em --epochs 100

Evaluation

For evaluation you can use the evaluate_models script:

python evaluate_models.p -h

usage: evaluate_models.py [-h] [--output OUTPUT] [--clusters CLUSTERS] [--amount AMOUNT] [--raw_net] [--debug] [--cuda] REAL FAKE MODELS

positional arguments:
  REAL                  Directory containing real data.
  FAKE                  Directory containing fake data.
  MODELS                Directory containing model checkpoints.

optional arguments:
  -h, --help            show this help message and exit
  --output OUTPUT, -o OUTPUT
                        Output file name.
  --clusters CLUSTERS, -k CLUSTERS
                        The amount of clusters to learn (default: 128).
  --amount AMOUNT, -a AMOUNT
                        Amount of files to load from each directory (default: None - all).
  --raw_net, -r         RawNet models?
  --debug, -d           Only use minimal amount of files?
  --cuda, -c            Use cuda?

Example

python evaluate_models.py /data/LJSpeech-1.1/wavs /data/generated_audio trained_models/lfcc/em

Make sure to move the out-of-distribution models to a seperate directory first!

Attribution

We provide a script to attribute the GMM models:

python attribute.py -h

usage: attribute.py [-h] [--clusters CLUSTERS] [--steps STEPS] [--blur] FILE REAL_MODEL FAKE_MODEL

positional arguments:
  FILE                  Audio sample to attribute.
  REAL_MODEL            Real model to attribute.
  FAKE_MODEL            Fake Model to attribute.

optional arguments:
  -h, --help            show this help message and exit
  --clusters CLUSTERS, -k CLUSTERS
                        The amount of clusters to learn (default: 128).
  --steps STEPS, -m STEPS
                        Amount of steps for integrated gradients.
  --blur, -b            Compute BlurIG instead.

Example

python attribute.py /data/LJSpeech-1.1/wavs/LJ008-0217.wav path/to/real/model.pth path/to/fake/model.pth

BibTeX

When you cite our work feel free to use the following bibtex entry:

@inproceedings{
  frank2021wavefake,
  title={{WaveFake: A Data Set to Facilitate Audio Deepfake Detection}},
  author={Joel Frank and Lea Sch{\"o}nherr},
  booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2021},
}

GitHub

View Github