README.md

metadata

license: apache-2.0
title: DeepFense
sdk: streamlit
emoji: 🛡️
colorFrom: blue
colorTo: indigo
pinned: true

🛡️ DeepFense: Modular Deepfake Audio Detection 🎙️

👋 Welcome to the official Hugging Face home of the DeepFense Framework!

We are the DeepFense Team. Our mission is to provide the research community with a robust, modular, and configuration-driven framework for building, training, and evaluating state-of-the-art audio deepfake detectors.

---

🚀 The Framework at a Glance

DeepFense simplifies audio forensics by decoupling frontends, backends, and loss functions. Whether you are a researcher or a developer, you can swap components via simple YAML configs—no code changes required.

• Modular Design: Mix-and-match components like WavLM, HuBERT, or EAT (Frontends) with AASIST, Nes2Net, or ECAPA-TDNN (Backends).
• Hugging Face Native: Access 455+ pretrained models and 12 benchmark datasets (ASVSpoof19, CompSpoof, etc.) directly through our CLI.
• Reproducible Research: Automatic tracking of EER, minDCF, and F1 metrics with built-in WandB logging.
• Rich Augmentations: Built-in RawBoost, RIR, Codec, and more to ensure model robustness.

---

🛠️ Quick Start

Get started with DeepFense in under 5 minutes:

# Install the framework
pip install deepfense

# Download a specific model & dataset
deepfense download dataset CompSpoof
deepfense download model ASV19_WavLM_Nes2Net_NoAug_Seed42