Weakly-supervised deepfake localization in diffusion-generated images

View on arXiv ← Back to list

Authors: Dragos Tantaru, Elisabeta Oneata, Dan Oneata

Published: 2023-11-08 10:27:36+00:00

AI Summary

This paper proposes a weakly-supervised framework for deepfake localization in images generated by diffusion models. It compares three localization methods (GradCAM, Patches, Attention) using a novel dataset with locally and fully manipulated images, analyzing the impact of supervision level, dataset, and generator.

Abstract

The remarkable generative capabilities of denoising diffusion models have raised new concerns regarding the authenticity of the images we see every day on the Internet. However, the vast majority of existing deepfake detection models are tested against previous generative approaches (e.g. GAN) and usually provide only a fake or real label per image. We believe a more informative output would be to augment the per-image label with a localization map indicating which regions of the input have been manipulated. To this end, we frame this task as a weakly-supervised localization problem and identify three main categories of methods (based on either explanations, local scores or attention), which we compare on an equal footing by using the Xception network as the common backbone architecture. We provide a careful analysis of all the main factors that parameterize the design space: choice of method, type of supervision, dataset and generator used in the creation of manipulated images; our study is enabled by constructing datasets in which only one of the components is varied. Our results show that weakly-supervised localization is attainable, with the best performing detection method (based on local scores) being less sensitive to the looser supervision than to the mismatch in terms of dataset or generator.

Key findings

The patch-based method significantly outperforms GradCAM and Attention in weakly-supervised settings. Weakly-supervised localization is feasible, with performance more sensitive to generator mismatch than supervision type. Generalization across different generators and datasets is challenging.

Approach

The paper frames deepfake localization as a weakly-supervised problem. It uses an Xception network backbone with three different methods to generate localization maps: GradCAM (explanations), Patches (local scores), and Attention (attention mechanisms). The performance is evaluated with image-level labels and, for comparison, ground truth masks.

Datasets

CelebA-HQ, FFHQ, datasets generated using P2 diffusion model, Repaint method with P2 and LDM models, LaMa, and Pluralistic inpainting methods. COCO Glide dataset used for out-of-domain evaluation.

Model(s)

Xception network as a backbone, GradCAM, Patches, Attention methods for localization.

Author countries

Romania

← Previous