Open Set Face Forgery Detection via Dual-Level Evidence Collection (OSFFD)

Official implementation of Open Set Face Forgery Detection via Dual-Level Evidence Collection (FG 2026).

This work proposes the Dual Level Evidential face forgery Detection (DLED) approach, a CLIP-based detector with two parallel branches (spatial / frequency) and fuses them via subjective-logic evidence combination.

Repository layout

DLED/
├── train.sh                       # unified launcher (replaces protocol1-8)
├── Dassl/                         # third-party Dassl toolbox
├── clip/                          # CLIP model
├── preprocessing/                 # scripts to build the DF40 JSON indices
└── training/
    ├── train.py                   # training entry-point
    ├── test.py                    # evaluation entry-point
    ├── config/
    │   ├── train_config.yaml
    │   ├── test_config.yaml
    │   └── detector/dled.yaml
    ├── dataset/                   # DeepfakeFrequencyDataset etc.
    ├── detectors/dled.py
    ├── trainer/trainer.py
    ├── loss/edl_loss.py
    ├── metrics/                   # registry + metric helpers
    └── optimizor/                 # SAM, LinearLR

Setup

pip install -r requirements.txt   # PyTorch, albumentations, timm, ftfy, regex, yacs, ...

The DF40 dataset must be available locally. Generate the JSON index files via the scripts under preprocessing/, then point the config at them.

Environment variables

Variable	Purpose
DLED_PRETRAINED_ROOT	Directory used to cache CLIP weights. Defaults to training/pretrained/clip.
DF40_DATA_ROOT	Root directory of the DF40 dataset on this machine.
DF40_OLD_ROOT	Root path embedded in the JSON index (if it differs from DF40_DATA_ROOT, paths are rewritten on the fly).
DLED_LOG_DIR	Output directory for training logs (default: ./logs).
DLED_LOCAL_RANK	CUDA device index (default: 0).
DLED_NUM_EPOCHS	Number of training epochs (default: 50).
DLED_TEST_FRE	Test frequency in epochs (default: 50).
DLED_MAX_SAMPLES_PER_CLS	Cap on the number of samples per training class (default: 13504).

Also edit training/config/train_config.yaml so that dataset_json_folder and log_dir point at the right locations on your machine.

Training

The single launcher train.sh supersedes the old train_protocol{1..8}.sh files. Pick the protocol you want to run:

# 3-class same-domain protocol with Entire-Face-Synthesis held out (paper Table 1)
./train.sh NOEFS

# 3-class same-domain protocol with Face-Swapping held out
./train.sh NOFS

# 3-class same-domain protocol with Face-Reenactment held out
./train.sh NOFR

# 4-class open-test protocol with Face-Editing held out (paper Table 2)
./train.sh open

Extra arguments after the protocol name are forwarded to training/train.py, e.g.:

./train.sh NOEFS --save_model --save_path ./saved_models/NOEFS

Evaluation

python3 training/test.py \
    --detector_path ./training/config/detector/dled.yaml \
    --train_dataset DF40_FF_train_NOEFS \
    --test_dataset  DF40_FF_test_NOFE_all \
    --num_classes 3 \
    --model_path  ./saved_models/NOEFS/<checkpoint>.pth

Citation

If you use this code, please cite the paper.

@article{cai2025open, title={Open Set Face Forgery Detection via Dual-Level Evidence Collection}, author={Cai, Zhongyi and Gernon, Bryce and Bao, Wentao and Li, Yifan and Wright, Matthew and Kong, Yu}, journal={arXiv preprint arXiv:2512.04331}, year={2025} }