Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fourier-Based GAN Fingerprint Detection using ResNet50

Sai Teja Erukude, Viswa Chaitanya Marella, Suhasnadh Reddy Veluru

TL;DR

The paper tackles the challenge of distinguishing StyleGAN-generated images from real ones for digital forensics. It proposes a Fourier-domain approach by applying a 2D DFT, followed by fftshift and log normalization, to create frequency-domain inputs that a ResNet50 classifier can learn from. The Fourier-based model achieves 92.82% accuracy and an AUC of 0.95 on a balanced test set, outperforming a spatial-domain baseline. This work demonstrates that GAN fingerprints exist in the Fourier spectrum and offers practical forensic relevance, while outlining future directions like cross-model generalization and alternative spectral representations.

Abstract

The rapid rise of photorealistic images produced from Generative Adversarial Networks (GANs) poses a serious challenge for image forensics and industrial systems requiring reliable content authenticity. This paper uses frequency-domain analysis combined with deep learning to solve the problem of distinguishing StyleGAN-generated images from real ones. Specifically, a two-dimensional Discrete Fourier Transform (2D DFT) was applied to transform images into the Fourier domain, where subtle periodic artifacts become detectable. A ResNet50 neural network is trained on these transformed images to differentiate between real and synthetic ones. The experiments demonstrate that the frequency-domain model achieves a 92.8 percent and an AUC of 0.95, significantly outperforming the equivalent model trained on raw spatial-domain images. These results indicate that the GAN-generated images have unique frequency-domain signatures or "fingerprints". The method proposed highlights the industrial potential of combining signal processing techniques and deep learning to enhance digital forensics and strengthen the trustworthiness of industrial AI systems.

Fourier-Based GAN Fingerprint Detection using ResNet50

TL;DR

The paper tackles the challenge of distinguishing StyleGAN-generated images from real ones for digital forensics. It proposes a Fourier-domain approach by applying a 2D DFT, followed by fftshift and log normalization, to create frequency-domain inputs that a ResNet50 classifier can learn from. The Fourier-based model achieves 92.82% accuracy and an AUC of 0.95 on a balanced test set, outperforming a spatial-domain baseline. This work demonstrates that GAN fingerprints exist in the Fourier spectrum and offers practical forensic relevance, while outlining future directions like cross-model generalization and alternative spectral representations.

Abstract

The rapid rise of photorealistic images produced from Generative Adversarial Networks (GANs) poses a serious challenge for image forensics and industrial systems requiring reliable content authenticity. This paper uses frequency-domain analysis combined with deep learning to solve the problem of distinguishing StyleGAN-generated images from real ones. Specifically, a two-dimensional Discrete Fourier Transform (2D DFT) was applied to transform images into the Fourier domain, where subtle periodic artifacts become detectable. A ResNet50 neural network is trained on these transformed images to differentiate between real and synthetic ones. The experiments demonstrate that the frequency-domain model achieves a 92.8 percent and an AUC of 0.95, significantly outperforming the equivalent model trained on raw spatial-domain images. These results indicate that the GAN-generated images have unique frequency-domain signatures or "fingerprints". The method proposed highlights the industrial potential of combining signal processing techniques and deep learning to enhance digital forensics and strengthen the trustworthiness of industrial AI systems.
Paper Structure (11 sections, 8 figures, 2 tables)

This paper contains 11 sections, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data and CNN Architecture
  3. Real Images Collection
  4. GAN-generated Images Collection
  5. Dataset Composition
  6. CNN Architecture
  7. Proposed work
  8. Fourier Transform
  9. Training Setup
  10. Result Analysis
  11. Conclusion

Figures (8)

  • Figure 1: A sample of Real images from the dataset, including human faces (from FFHQ) and cat images (from the Cats vs. Dogs dataset). These images are the true class for the binary classification task. They also provide variety in terms of age, appearance, and background.
  • Figure 2:
  • Figure 3: High-level architecture of the ResNet50 deep neural network used for both spatial-domain and frequency-domain classification tasks.
  • Figure 4: Flowchart illustrating the methodology for Fourier-based GAN fingerprint extraction and detection.
  • Figure 5: Example showing a real cat image in (left) the spatial domain (raw pixel intensity) and (right) the frequency domain (magnitude of the 2D DFT). The frequency domain can delineate energy in specific frequency bands, which may exhibit qualitatively different structures or randomness in comparison to the generated images from GANs, which is mostly relevant for detection.
  • ...and 3 more figures