Skip to content

README.md

Latest commit

 

History

History
56 lines (41 loc) · 2.51 KB

README.md

File metadata and controls

56 lines (41 loc) · 2.51 KB
Raw

CMF-Lite

CMF-Lite is a lightweight cross-cue fusion detector for deepfake video detection. The model combines an EfficientNet-B0 RGB branch, a fixed Laplacian high-frequency residual branch, and a frozen CLIP ViT-B/32 semantic branch through a compact fusion head. Video-level predictions are produced by averaging sampled frame probabilities.

CMF-Lite Model

Protocol

The default study trains on FaceForensics++ C23 using the official source-video split and the four original manipulation folders: Deepfakes, Face2Face, FaceSwap, and NeuralTextures, together with original real videos. The manifest builder audits source identifiers so that train, validation, and test source IDs do not overlap. Celeb-DF v2 and WildDeepfake are used as external evaluations.

Main scripts

export PYTHONPATH=$PWD/src
python scripts/00_validate_environment.py --config configs/study.yaml
python scripts/00_fetch_official_ffpp_splits.py --config configs/study.yaml
python scripts/01_build_manifests.py --config configs/study.yaml
bash scripts/run_full_pipeline.sh

The full pipeline builds manifests, trains/evaluates the main model, runs cue ablations, evaluates robustness, generates qualitative error views, profiles resources, computes bootstrap statistics, and writes tables/figures under outputs/.

Generated artifacts

Key manuscript-facing outputs include:

  • outputs/reports/ffpp_protocol_summary.json
  • outputs/reports/ffpp_split_overlap.csv
  • outputs/reports/dataset_split_counts.csv
  • outputs/tables/ablation_cue_results.csv and .tex
  • outputs/tables/resource_profile.csv and .tex
  • outputs/tables/bootstrap_auc_ci.csv
  • outputs/figures/compression_robustness_video_auroc.png
  • outputs/figures/frame_sampling_video_auroc.png
  • outputs/figures/threshold_sensitivity_video_f1.png
  • outputs/figures/resource_tradeoff_auc_latency.png
  • outputs/figures/resource_tradeoff_auc_params.png

Notes

Large datasets, extracted crops, temporary frames, model checkpoints, and cached tensors are intentionally not included. Recreate them locally from the scripts above.

Citation

If you use this codebase, please cite the associated manuscript:

@inproceedings{oladele2026cmf_lite,
  title     = {CMF-Lite: Lightweight Cross-Cue Fusion for Generalizable Deepfake Video Detection},
  author    = {Oladele, Daniel Ayo and Sibiya, Malusi},
  booktitle = {Proceedings of MIWAI 2026},
  series    = {Lecture Notes in Artificial Intelligence},
  publisher = {Springer},
  year      = {2026}
}