---
license: apache-2.0
pipeline_tag: video-classification
---

# Referee: Reference-aware Audiovisual Deepfake Detection

This repository contains the `Referee` model, presented in the paper [Referee: Reference-aware Audiovisual Deepfake Detection](https://huggingface.co/papers/2510.27475).

Code: [https://github.com/ewha-mmai/referee](https://github.com/ewha-mmai/referee)

## Abstract
<img src="https://github.com/ewha-mmai/referee/raw/main/referee.png" alt="Referee Architecture" width="900"/>

Since deepfakes generated by advanced generative models have rapidly posed serious threats, existing audiovisual deepfake detection approaches struggle to generalize to unseen forgeries. We propose a novel reference-aware audiovisual deepfake detection method, called *Referee*. Speaker-specific cues from only one-shot examples are leveraged to detect manipulations beyond spatiotemporal artifacts. By matching and aligning identity-related queries from reference and target content into cross-modal features, Referee jointly reasons about audiovisual synchrony and identity consistency. Extensive experiments on FakeAVCeleb, FaceForensics++, and KoDF demonstrate that Referee achieves state-of-the-art performance on cross-dataset and cross-language evaluation protocols. Experimental results highlight the importance of cross-modal identity verification for future deepfake detection.

## Requirements
### Environment
To train or evaluate Referee, you must first set up the environment:

```bash
conda create -n referee python=3.8.16
conda activate referee
pip install -r requirements.txt
```

### Dataset
For training and evaluation, the dataset should be prepared following the specified format.
An example dataset structure is provided in the [GitHub repository's `data/`](https://github.com/ewha-mmai/referee/tree/main/data).

### Pretrained Checkpoints
This project requires pretrained checkpoints to run training, evaluation, or fine-tuning.

-   **Training from Scratch**  
    To train the model from scratch, download the Synchformer checkpoint trained on **LRS3** from the [link](https://github.com/v-iashin/Synchformer) and place it in the `model/pretrained/` directory.

-   **Evaluation or Fine-tuning Referee**  
    To evaluate or fine-tune **Referee**, download the provided checkpoint from the [link](https://huggingface.co/eunsanglee/Referee/tree/main) and put it into the `model/pretrained/` directory.

## Train
To train Referee, you can use the provided `train.sh`. Some training-specific settings, such as the number of epochs, starting epoch, and training dataset, are set directly in `train.sh`.

You can change most training parameters in the config file, `configs/pair_sync.yaml`. For example, you can adjust the learning rate, batch size, number of layers, etc.

Once you have set all parameters as desired, you can start training Referee using:

```bash
sh scripts/train.sh
```

## Evaluation
To evaluate Referee, you can use the provided `test.sh`. Some evaluation-specific settings, such as the model path and test dataset, are set directly in `test.sh`.

You can change most evaluation parameters in the config file, `configs/pair_sync.yaml`. For example, you can adjust the number of layers, the number of identity queries, etc.

Once you have set all parameters as desired, you can start evaluating Referee using:

```bash
sh scripts/test.sh
```

## Acknowledgement
This project heavily references the implementation of [SynchFormer](https://github.com/v-iashin/Synchformer).

We thank the authors for making their code publicly available.

## Citation
If you find our work helpful or inspiring, please feel free to cite it:

```bibtex
@article{boo2025referee,
  title={Referee: Reference-aware Audiovisual Deepfake Detection},
  author={Boo, Hyemin and Lee, Eunsang and Lee, Jiyoung},
  journal={arXiv preprint arXiv:2510.27475},
  year={2025}
}
```