DeepFense: A Unified, Modular, and Extensible Framework for Robust Deepfake Audio Detection

A research team led by Yassine El Kheir from TU Berlin, in collaboration with eight other authors, has released DeepFense, a comprehensive open-source toolkit built in PyTorch. This framework directly addresses a critical reproducibility crisis in the speech deepfake detection field by providing a unified, modular, and extensible codebase. It integrates the latest model architectures, loss functions, and data augmentation pipelines, alongside a library of over 100 ready-to-use training and evaluation recipes. This standardization allows for the first large-scale, apples-to-apples comparison of detection methodologies.

Using DeepFense, the team conducted an unprecedented evaluation of more than 400 distinct detection models. Their key finding is that the choice of pre-trained front-end audio feature extractor—such as Wav2Vec2 or HuBERT—is the dominant factor, accounting for roughly 70% of overall performance variance across datasets. While curated training data improves a model's ability to generalize to unseen audio, the front-end selection is paramount. Crucially, the analysis exposed severe, previously unquantified biases in even the highest-performing models, which show significantly degraded accuracy on lower-quality audio, female speakers, and non-English languages.

The DeepFense toolkit is designed to move the field from academic benchmarks to real-world deployment. By providing the tools to systematically analyze and mitigate these performance gaps and biases, developers can build more robust and equitable detection systems. The framework's modular design also allows researchers to rapidly prototype new detection strategies, accelerating innovation against increasingly sophisticated audio deepfakes generated by models like Vall-E and AudioLM. The complete code and recipes are publicly available to foster collaboration and transparency in this critical security domain.