DeepFense-CodecFake-W2V2-MLP: Speech Anti-Spoofing Arena wrapper

Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.

DeepFense (Kheir et al., 2026) CodecFake recipe: a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end (last transformer layer, features_only) followed by an MLP back-end (Linear 1024->128, LayerNorm, ReLU, mean pooling over time) and a 2-way linear CrossEntropy head. Trained on the Codecfake corpus (Xie et al., AISHELL-3/VCTK-based codec re-synthesis) without data augmentation, seed 42; the released best_model.pth (epoch 2) as is. Input as the DeepFense eval pipeline: raw float waveform (no normalisation), first 64000 samples (4 s), shorter clips repeat-padded. score = logit(bonafide) - logit(spoof); FP32.

• Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the meta.yaml notes.
• Params: 317.522306 M. Trained on Codecfake (Xie et al.) train, no augmentation.
• Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; the other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; FP32 on the RTX 5060 Ti), which matched PyTorch on the full J-SPAW_LA set.

Arena results

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.

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}
}