LFM (BMVC 2025)

Official PyTorch implementation of LFM (BMVC 2025). This repository contains training and evaluation code for our paper "A Novel Local Focusing Mechanism for Deepfake Detection Generalization" accepted at BMVC 2025.

📦 Installation

Prerequisites

NVIDIA GPU + CUDA cuDNN
Python 3.10
PyTorch 2.6+

Setup Environment

# Clone repository
git clone https://github.com/lmlpy/LFM.git
cd LFM

# Create and activate conda environment
conda create -n LFM python=3.10.16
conda activate LFM

# Install dependencies
pip install -r requirements.txt

Data preparation

Getting the data

	paper	Url
Train set	CNNDetection CVPR2020	Baidudrive
Val set	CNNDetection CVPR2020	Baidudrive
Table1 Test	CNNDetection CVPR2020	Baidudrive
Table2 Test	FreqNet AAAI2024	googledrive
Table3 Test	DIRE ICCV2023	googledrive
Table4 Test	UniversalFakeDetect CVPR2023	googledrive
Table5 Test	NPR CVPR2024	googledrive

Data structure

Click to expand the folder tree structure.

datasets
|-- ForenSynths_train_val
|   |-- train
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   `-- horse
|   `-- val
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   `-- horse
|   |-- test
|       |-- biggan
|       |-- cyclegan
|       |-- deepfake
|       |-- gaugan
|       |-- progan
|       |-- stargan
|       |-- stylegan
|       `-- stylegan2
`-- Generalization_Test
    |-- ForenSynths_test       # Table1
    |   |-- biggan
    |   |-- cyclegan
    |   |-- deepfake
    |   |-- gaugan
    |   |-- progan
    |   |-- stargan
    |   |-- stylegan
    |   `-- stylegan2
    |-- GANGen-Detection     # Table2
    |   |-- AttGAN
    |   |-- BEGAN
    |   |-- CramerGAN
    |   |-- InfoMaxGAN
    |   |-- MMDGAN
    |   |-- RelGAN
    |   |-- S3GAN
    |   |-- SNGAN
    |   `-- STGAN
    |-- DiffusionForensics  # Table3
    |   |-- adm
    |   |-- ddpm
    |   |-- iddpm
    |   |-- ldm
    |   |-- pndm
    |   |-- sdv1_new
    |   |-- sdv2
    |   `-- vqdiffusion
    `-- UniversalFakeDetect # Table4
    |   |-- dalle
    |   |-- glide_100_10
    |   |-- glide_100_27
    |   |-- glide_50_27
    |   |-- guided          # Also known as ADM.
    |   |-- ldm_100
    |   |-- ldm_200
    |   `-- ldm_200_cfg
    |-- Diffusion1kStep     # Table5
        |-- DALLE
        |-- ddpm
        |-- guided-diffusion    # Also known as ADM.
        |-- improved-diffusion  # Also known as IDDPM.
        `-- midjourney

Training

python train.py --name lfm --dataroot ./datasets/ForenSynths_train_val --classes car,cat,chair,horse \
       --batch_size 32 --delr_freq 10 --lr 0.0001 --niter 50

Testing

python test.py --model_path ./weights/bs_model.pth --batch_size 32

📌 LFM Visualization

Acknowledgments

This repository borrows partially from the NPR.

📜 Citation

If you find this repository useful for your research, please consider citing:

@misc{li2025novellocalfocusingmechanism,
      title={A Novel Local Focusing Mechanism for Deepfake Detection Generalization}, 
      author={Mingliang Li and Lin Yuanbo Wu and Changhong Liu and Hanxi Li},
      year={2025},
      eprint={2508.17029},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.17029}, 
}