Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Parents and Children: Distinguishing Multimodal DeepFakes from Natural Images

Roberto Amoroso, Davide Morelli, Marcella Cornia, Lorenzo Baraldi, Alberto Del Bimbo, Rita Cucchiara

TL;DR

The paper tackles the rising challenge of diffusion-model deepfakes by introducing a multimodal detection framework that leverages both perceptual cues and semantic information. It formulates a two-branch disentanglement approach with separate style and semantic projections, trained via supervised contrastive loss to distinguish real versus fake images even when low-level cues are removed. A key contribution is the COCOFake dataset, containing over 1.2 million images generated from COCO captions using Stable Diffusion v1.4 and v2.0, enabling large-scale evaluation of semantic-consistent forgery. Empirical results show that contrastive, multimodal features yield high detection performance, with the style space providing strong discrimination and the semantic space offering insight into semantic preservation and robustness across transformations and model variants. This work advances practical deepfake detection by combining semantic disentanglement with a large, semantically structured dataset, supporting robust identification of diffusion-model forgeries in real-world settings.

Abstract

Recent advancements in diffusion models have enabled the generation of realistic deepfakes from textual prompts in natural language. While these models have numerous benefits across various sectors, they have also raised concerns about the potential misuse of fake images and cast new pressures on fake image detection. In this work, we pioneer a systematic study on deepfake detection generated by state-of-the-art diffusion models. Firstly, we conduct a comprehensive analysis of the performance of contrastive and classification-based visual features, respectively extracted from CLIP-based models and ResNet or ViT-based architectures trained on image classification datasets. Our results demonstrate that fake images share common low-level cues, which render them easily recognizable. Further, we devise a multimodal setting wherein fake images are synthesized by different textual captions, which are used as seeds for a generator. Under this setting, we quantify the performance of fake detection strategies and introduce a contrastive-based disentangling method that lets us analyze the role of the semantics of textual descriptions and low-level perceptual cues. Finally, we release a new dataset, called COCOFake, containing about 1.2M images generated from the original COCO image-caption pairs using two recent text-to-image diffusion models, namely Stable Diffusion v1.4 and v2.0.

Parents and Children: Distinguishing Multimodal DeepFakes from Natural Images

TL;DR

The paper tackles the rising challenge of diffusion-model deepfakes by introducing a multimodal detection framework that leverages both perceptual cues and semantic information. It formulates a two-branch disentanglement approach with separate style and semantic projections, trained via supervised contrastive loss to distinguish real versus fake images even when low-level cues are removed. A key contribution is the COCOFake dataset, containing over 1.2 million images generated from COCO captions using Stable Diffusion v1.4 and v2.0, enabling large-scale evaluation of semantic-consistent forgery. Empirical results show that contrastive, multimodal features yield high detection performance, with the style space providing strong discrimination and the semantic space offering insight into semantic preservation and robustness across transformations and model variants. This work advances practical deepfake detection by combining semantic disentanglement with a large, semantically structured dataset, supporting robust identification of diffusion-model forgeries in real-world settings.

Abstract

Recent advancements in diffusion models have enabled the generation of realistic deepfakes from textual prompts in natural language. While these models have numerous benefits across various sectors, they have also raised concerns about the potential misuse of fake images and cast new pressures on fake image detection. In this work, we pioneer a systematic study on deepfake detection generated by state-of-the-art diffusion models. Firstly, we conduct a comprehensive analysis of the performance of contrastive and classification-based visual features, respectively extracted from CLIP-based models and ResNet or ViT-based architectures trained on image classification datasets. Our results demonstrate that fake images share common low-level cues, which render them easily recognizable. Further, we devise a multimodal setting wherein fake images are synthesized by different textual captions, which are used as seeds for a generator. Under this setting, we quantify the performance of fake detection strategies and introduce a contrastive-based disentangling method that lets us analyze the role of the semantics of textual descriptions and low-level perceptual cues. Finally, we release a new dataset, called COCOFake, containing about 1.2M images generated from the original COCO image-caption pairs using two recent text-to-image diffusion models, namely Stable Diffusion v1.4 and v2.0.
Paper Structure (16 sections, 5 equations, 7 figures, 7 tables)

This paper contains 16 sections, 5 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Notation and Preliminaries
  5. Learning to Discriminate Real and Fake images
  6. Semantic Preservation Analysis
  7. Disentangling Semantics and Style
  8. The COCOFake Dataset for Multimodal Deepfake Recognition
  9. Experimental Evaluation
  10. Implementation Details
  11. Metrics
  12. Performance of Visual Features
  13. Semantic-Style Disentangling Results
  14. Robustness Analysis to Image Transformations
  15. Comparison with Other Methods
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Overview of our multimodal deepfakes detection setting, in which five subsets of the semantics contained in a given image are employed to generate as many fake images.
  • Figure 2: Schema of our approach for disentangling semantics and style for deepfake detection.
  • Figure 3: Sample images from COCOFake. The leftmost column shows the original (real) image, while the remaining ones show fake images generated by Stable Diffusion v1.4 from each of the five COCO captions.
  • Figure 4: Less realistic images from COCOFake. The leftmost column shows the original (real) image, while the remaining ones show fake images generated by Stable Diffusion v1.4 from each of the five COCO captions.
  • Figure 5: Sample misclassification errors on both real (left) and fake (right) images, using OpenCLIP ViT-B/32 trained on LAION-2B as the visual encoder.
  • ...and 2 more figures