AOCloss: Adaptive Centroid Shift Loss for Audio Deepfake Detection

📖 Overview

This repository provides an implementation of the Adaptive Centroid Shift Loss (AOCloss) method for Audio Deepfake Detection, as described in the corresponding research paper. The approach employs a one-class learning framework that continuously adapts a centroid to represent bonafide audio embeddings while maximizing the distance of spoof embeddings.

🛠️ Usage

1. Initialization

from AOC_loss import AOCloss

# Initialize with desired embedding dimension
criterion = AOCloss(embedding_dim=512)

2. Using the AOC loss

loss = criterion(embeddings, labels)

• embeddings: Tensor of shape (batch_size, embedding_dim).
• labels: Binary tensor where 0 represents bonafide samples and 1 represents spoof samples.

3. Centroid Update

The centroid is automatically updated during the forward pass.

criterion.update_centroid(bonafide_embeddings)

✅ Requirements

• PyTorch >= 1.10
• Python >= 3.8

❓ Troubleshooting

• ValueError: Centroid has not been initialized:

  • Ensure the batch contains bonafide samples.

• Negative Loss:

  • This is expected due to the range of cosine similarity between -1 and 1.

📄 Citation

This github implementation based on the following paper:

@inproceedings{kim24b_interspeech,
  title     = {One-class learning with adaptive centroid shift for audio deepfake detection},
  author    = {Hyun Myung Kim and Kangwook Jang and Hoirin Kim},
  year      = {2024},
  booktitle = {Interspeech 2024},
  pages     = {4853--4857},
  doi       = {10.21437/Interspeech.2024-177},
  issn      = {2958-1796},
}

🙋‍♂️ Support

For issues or questions, feel free to open an issue in the repository.

📢 Acknowledgments

• This implementation is inspired by research paper.