Practical Manipulation Model for Robust Deepfake Detection

View on arXiv ← Back to list

Authors: Benedikt Hopf, Radu Timofte

Published: 2025-06-05 15:06:16+00:00

AI Summary

This paper introduces a Practical Manipulation Model (PMM) for robust deepfake detection, improving upon existing methods' vulnerability to non-ideal conditions. PMM enhances the diversity of pseudo-fake training data by incorporating Poisson blending, diverse masks, generator artifacts, and image degradations, leading to significantly improved robustness and benchmark performance.

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. Code can be found at https://github.com/BenediktHopf/PMM

Key findings

PMM significantly improves the Area Under the Curve (AUC) on DFDC and DFDCP datasets compared to state-of-the-art models like LAA. The method shows improved robustness against common image degradations like noise and blur. The improvements are consistent across different backbone architectures, demonstrating the generality of the PMM approach.

Approach

The authors improve deepfake detection robustness by augmenting the training data with a Practical Manipulation Model (PMM). PMM generates more realistic pseudo-fakes using Poisson blending, diverse masks, generator artifacts, and a range of image degradations. This approach enhances the model's generalization ability and resistance to real-world image imperfections.

Datasets

FaceForensics++, DFDC, DFDCP, Celeb-DF-v2, DF40

Model(s)

Localized Artifact Attention (LAA), EfficientNet-b4, Xception

Author countries

Germany

← Previous