Comprehensive Layer-wise Analysis of SSL Models for Audio Deepfake Detection

This repository contains the official implementation of the paper:
"Comprehensive Layer-wise Analysis of SSL Models for Audio Deepfake Detection" (NAACL Findings 2025).
[Paper]

---

Abstract

This work conducts a systematic layer-wise analysis of self-supervised learning (SSL) models (e.g., Wav2Vec2, Hubert, WavLM) for detecting audio deepfakes across multilingual datasets (English, Chinese, Spanish), partial/song/scene-based deepfakes, and varying acoustic conditions. Key findings include:

• Lower layers (1-6 for small models, 1-12 for large models) provide the most discriminative features for detection.
• Reduced-layer models achieve competitive performance while lowering computational costs.
• Results generalize across languages (En, Zh, Es) and deepfake scenarios (full/partial/song/scene).

Weights Analysis

---

Installation

1. Clone the repository:

   git clone https://github.com/Yaselley/SSL_Layerwise_Deepfake
   cd SSL_Layerwise_Deepfake

Install dependencies

To install the required dependencies, run the following command:

pip install -r requirements.txt

Datasets

Publicly available datasets used in this work:

• ASVspoof 2019 (LA19): Link
• ASVspoof 2021 (LA21/DF21): Link
• ADD23: Link
• HABLA: Link
• PartialSpoof: Link
• SceneFake: Link
• CtrSVDD: Link

Usage

Replicating Experiments

Each dataset has a dedicated script. Example for ASVspoof 2019 (LA19):

python main.py \
  --seed 42 \
  --model w2v \          # Options: w2v, hubert, wavlm
  --small \              # Use small SSL model variant
  --n_layers 6 \          # Number of transformer layers (e.g., 6 for small models)
  --back FFN             # Backend classifier: FFN or AASIST

Other Scripts

• Scene-based deepfakes: main_sceneFake.py
• Spanish (HABLA): main_spanish.py
• Partial deepfakes: main_partial.py

Analysis

• Heatmaps (layer-wise weights) and tables are generated in the analysis/ folder.

Conclusion

• Lower SSL layers are critical for detecting artifacts in synthetic audio.
• Reducing layers (e.g., 4-6 for small models) achieves competitive performance while improving inference speed by 2-3×.
• Models generalize across languages and deepfake types, offering practical solutions for real-world deployment.

Citation

@misc{kheir2025comprehensivelayerwiseanalysisssl,
  title={Comprehensive Layer-wise Analysis of SSL Models for Audio Deepfake Detection}, 
  author={Yassine El Kheir and Youness Samih and Suraj Maharjan and Tim Polzehl and Sebastian Möller},
  year={2025},
  eprint={2502.03559},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2502.03559}, 
}