Partial deepfake / PartialSpoof: This README and the default master branch cover clip-level detection. For per-frame labels, TemporalDetector, Range EER, PartialSpoof metrics, and related configs, switch to the deepfense-partial branch:

git fetch origin
git checkout deepfense-partial

---

What is DeepFense?

DeepFense lets you build deepfake audio detectors by combining frontends (pretrained feature extractors), backends (classifiers), and loss functions — all defined in a single YAML config. No code changes needed to run new experiments.

The diagram above shows both clip-level and partial deepfake pipelines. On master, training and docs focus on clip-level detection (one score per utterance). For partial deepfake, use git checkout deepfense-partial (see note above).

Raw Audio --> Frontend (Wav2Vec2, WavLM, HuBERT, ...) --> Backend (AASIST, MLP, ...) --> Loss --> Score

---

Install

conda create -n deepfense python=3.10
conda activate deepfense
pip install deepfense

Or install from source (for development):

conda create -n deepfense python=3.10
conda activate deepfense
git clone https://github.com/Yaselley/deepfense-framework
cd deepfense-framework
pip install -e .

---

Quick Start

1. Generate dummy test data

python tests/create_samples.py

2. Train

python train.py --config deepfense/config/train.yaml

3. Test

python test.py \
    --config deepfense/config/train.yaml \
    --checkpoint outputs/Wav2Vec2_Nes2Net_Example_*/best_model.pth

4. Multi-GPU Training

DeepFense supports multi-GPU training out of the box via PyTorch DDP. Just use torchrun:

# 2 GPUs on a single node
torchrun --nproc_per_node=2 train.py --config deepfense/config/train.yaml

# 4 GPUs
torchrun --nproc_per_node=4 train.py --config deepfense/config/train.yaml

No config changes required -- DDP is detected automatically. Checkpoints, logs, and evaluation run on rank 0 only. The saved checkpoints are identical to single-GPU ones and can be loaded without any DDP-specific handling.

5. Use real data

Create a Parquet file with columns ID, path, label ("bonafide" / "spoof"), then update the config:

data:
  train:
    parquet_files: ["/path/to/train.parquet"]
  val:
    parquet_files: ["/path/to/val.parquet"]

---

How Configuration Works

Everything is controlled by a single YAML file. Here is the anatomy:

# ---------- experiment ----------
exp_name: "my_experiment"
output_dir: "./outputs/"
seed: 42

# ---------- data ----------
data:
  sampling_rate: 16000
  label_map: {"bonafide": 1, "spoof": 0}
  train:
    parquet_files: ["train.parquet"]
    batch_size: 32
    base_transform:
      - type: "pad"
        max_len: 64600          # ~4 sec at 16 kHz
    augment_transform:          # training only
      - type: "rawboost"
        noise_ratio: 0.4
  val:
    parquet_files: ["val.parquet"]
    batch_size: 64
    base_transform:
      - type: "pad"
        max_len: 64600

# ---------- model ----------
model:
  type: "StandardDetector"
  frontend:
    type: "wav2vec2"                    # or wavlm, hubert, mert, eat
    args:
      source: "huggingface"             # or "fairseq" for local .pt files
      ckpt_path: "facebook/wav2vec2-xls-r-300m"
      freeze: True
  backend:
    type: "AASIST"                      # or MLP, Nes2Net, ECAPA_TDNN, RawNet2
    args:
      input_dim: 1024                   # must match frontend output dim
  loss:
    - type: "OCSoftmax"                 # or CrossEntropy, AMSoftmax, ASoftmax
      weight: 1.0
      embedding_dim: 32                 # must match backend output dim

# ---------- training ----------
training:
  epochs: 50
  device: "cuda"
  optimizer:
    type: "adam"
    lr: 0.0001
  scheduler:
    type: "cosine_annealing"
    T_max: 50
  monitor_metric: "EER"
  monitor_mode: "min"
  metrics:
    EER: {}
    ACC: {}
    minDCF: {Pspoof: 0.05}

See the Full Tutorial for a detailed walkthrough of every parameter.

---

Available Components

Category	Options
Frontends	Wav2Vec2, WavLM, HuBERT, MERT, EAT
Backends	AASIST, ECAPA-TDNN, Nes2Net, RawNet2, MLP, TCM
Losses	CrossEntropy, OC-Softmax, AM-Softmax, A-Softmax
Augmentations	RawBoost, RIR, Codec, AdditiveNoise, SpeedPerturb, AddBabble, DropChunk, DropFreq
Metrics	EER, minDCF, actDCF, ACC, F1

List them from the CLI:

deepfense list
deepfense list --component-type backends

---

Pretrained Models & Datasets (HuggingFace Hub)

DeepFense publishes 455+ pretrained models and 12 datasets at huggingface.co/DeepFense.

# See what's available
deepfense download list-datasets
deepfense download list-models --filter WavLM

# Download a dataset (parquet files)
deepfense download dataset CompSpoof

# Download a pretrained model (checkpoint + config)
deepfense download model ASV19_WavLM_Nes2Net_NoAug_Seed42

# Test the downloaded model
python test.py \
    --config models/ASV19_WavLM_Nes2Net_NoAug_Seed42/config.yaml \
    --checkpoint models/ASV19_WavLM_Nes2Net_NoAug_Seed42/best_model.pth

Or in Python:

from deepfense.hub import download_dataset, download_model

parquets = download_dataset("CompSpoof")           # returns list of local paths
files    = download_model("ASV19_WavLM_Nes2Net_NoAug_Seed42")  # returns {"checkpoint": ..., "config": ...}

See the HuggingFace Hub Guide for full workflows (training, evaluation, inference).

---

Adding Your Own Component

Every component type follows the same pattern:

1. Create a file (e.g. deepfense/models/backends/my_backend.py)
2. Decorate with the registry:

   from deepfense.utils.registry import register_backend
   from deepfense.models.base_model import BaseBackend
   
   @register_backend("MyBackend")
   class MyBackend(BaseBackend):
       def __init__(self, config):
           super().__init__()
           # ...
   
       def forward(self, x):
           # ...
   

3. Import it in the package __init__.py
4. Use it in your config:

   backend:
     type: "MyBackend"
     args: { ... }
   

The same pattern applies to frontends, losses, augmentations, datasets, optimizers, and metrics. See the user guides for detailed walkthroughs.

---

Project Structure

deepfense/
├── cli/             # CLI commands (train, test, list)
├── config/          # YAML configs + parquet generators
├── data/            # Dataset loading + transforms/augmentations
├── models/
│   ├── frontends/   # Wav2Vec2, WavLM, HuBERT, MERT, EAT
│   ├── backends/    # AASIST, ECAPA-TDNN, Nes2Net, MLP, ...
│   ├── losses/      # OC-Softmax, AM-Softmax, CrossEntropy, ...
│   └── modules/     # Shared layers (pooling, conformer, fairseq_local)
├── training/        # Trainer, evaluator, metrics, seed
└── utils/           # Registry, visualization

---

Documentation

Guide	Description
Installation	Setup instructions
Quick Start	First model in 5 minutes
Full Tutorial	Every config option explained
Architecture	How DeepFense works internally
Configuration Reference	All YAML parameters
Library Usage	Use DeepFense as a Python library
HuggingFace Hub	Download datasets & pretrained models
CLI Reference	CLI commands
Components	Frontend, backend, loss, augmentation reference
User Guides	Adding custom components, training workflows

---

Citation

@article{kheir2026deepfense,
  title={DeepFense: A Unified, Modular, and Extensible Framework for Robust Deepfake Audio Detection},
  author={Kheir, Yassine El and Das, Arnab and Xiao, Yixuan and Wang, Xin and Kallel, Feidi and Erdogan, Enes Erdem and Vu, Ngoc Thang and Polzehl, Tim and Moeller, Sebastian},
  journal={arXiv preprint arXiv:2604.08450},
  year={2026}
}

License

Apache 2.0 -- see LICENSE for details.