tags:
- deepfake
- detection
- robustness
Model Card for PMM
This is the model card for the PMM model from Practical Manipulation Model for Robust Deepfake Detection
Abstract
Modern deepfake detection models have achieved strong performance even on the challenging cross-dataset task. However, detection performance under non-ideal conditions remains very unstable, limiting success on some benchmark datasets and making it easy to circumvent detection. Inspired by the move to a more real-world degradation model in the area of image super-resolution, we have developed a Practical Manipulation Model (PMM) that covers a larger set of possible forgeries. We extend the space of pseudo-fakes by using Poisson blending, more diverse masks, generator artifacts, and distractors. Additionally, we improve the detectors' generality and robustness by adding strong degradations to the training images. We demonstrate that these changes not only significantly enhance the model's robustness to common image degradations but also improve performance on standard benchmark datasets. Specifically, we show clear increases of 3.51% and 6.21% AUC on the DFDC and DFDCP datasets, respectively, over the s-o-t-a LAA backbone. Furthermore, we highlight the lack of robustness in previous detectors and our improvements in this regard.
Model Details
Model Description
- Developed by: Benedikt Hopf and Radu Timofte
- License: See github repository: https://github.com/benedikthopf/PMM
- Finetuned from model [optional]: LAA-Net: https://github.com/10Ring/LAA-Net
Model Sources
- Repository: https://github.com/benedikthopf/PMM
- Paper: Practical Manipulation Model for Robust Deepfake Detection
Uses
This model is a research project, indended to improve the robustness of deepfake detectors.
Direct Use
The model has been trained and tested on/for the FaceForensics++ benchmark with cross dataset testing on CDFv2, DFDC and DFDCP.
Downstream Use
Detection of newer face-swapping methods.
Out-of-Scope Use
The model is not built to detect fully synthetic images. Furthermore, IT IS NOT PERFECT AND MAY ALWAYS MAKE MISTAKES!
Bias, Risks, and Limitations
It is only designed and trained for detecting face-swaps/reenactment, not fully synthetic images.
We know that it performs slightly worse on clean data compared to the base model.
Citation (Bibtex)
@InProceedings{hopf2025practicalmanipulationmodelrobust,
author = {Hopf, Benedikt and Timofte, Radu},
title = {Practical Manipulation Model for Robust Deepfake Detection},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
month = {October},
year = {2025},
pages = {5675-5684}
}