Wavelet-based GAN Fingerprint Detection using ResNet50

Authors: Sai Teja Erukude, Suhasnadh Reddy Veluru, Viswa Chaitanya Marella

Published: 2025-10-21 22:40:16+00:00

Comment: 6 pages; Published in IEEE

AI Summary

This research introduces a wavelet-based method for detecting images generated by StyleGANs, utilizing discrete wavelet transform (DWT) preprocessing followed by a ResNet50 classifier. By transforming images into multi-resolution representations using Haar and Daubechies wavelet filters, the method capitalizes on subtle GAN-generated artifacts in the frequency domain. This approach significantly outperforms traditional spatial-domain detection, demonstrating the effectiveness of wavelet-domain analysis for identifying GAN fingerprints.

Abstract

Identifying images generated by Generative Adversarial Networks (GANs) has become a significant challenge in digital image forensics. This research presents a wavelet-based detection method that uses discrete wavelet transform (DWT) preprocessing and a ResNet50 classification layer to differentiate the StyleGAN-generated images from real ones. Haar and Daubechies wavelet filters are applied to convert the input images into multi-resolution representations, which will then be fed to a ResNet50 network for classification, capitalizing on subtle artifacts left by the generative process. Moreover, the wavelet-based models are compared to an identical ResNet50 model trained on spatial data. The Haar and Daubechies preprocessed models achieved a greater accuracy of 93.8 percent and 95.1 percent, much higher than the model developed in the spatial domain (accuracy rate of 81.5 percent). The Daubechies-based model outperforms Haar, showing that adding layers of descriptive frequency patterns can lead to even greater distinguishing power. These results indicate that the GAN-generated images have unique wavelet-domain artifacts or fingerprints. The method proposed illustrates the effectiveness of wavelet-domain analysis to detect GAN images and emphasizes the potential of further developing the capabilities of future deepfake detection systems.

Key findings

The Daubechies-based model achieved the highest accuracy of 95.1% and an AUC of 0.97, followed by the Haar wavelet model with 93.8% accuracy and 0.96 AUC. Both wavelet-based models significantly outperformed the ResNet50 model trained on spatial data, which only achieved 81.5% accuracy and 0.85 AUC. This indicates that GAN-generated images possess unique wavelet-domain artifacts or 'fingerprints' that are more effectively detected through frequency analysis.

Approach

The proposed method preprocesses input images using Discrete Wavelet Transform (DWT) with either Haar or Daubechies wavelet filters, converting them into multi-resolution frequency representations. These transformed images are then fed into a ResNet50 convolutional neural network for binary classification, distinguishing between real and StyleGAN-generated images.

Datasets

Flickr-Faces-HQ (FFHQ), Cats vs Dogs dataset (Kaggle), publicly available collections of StyleGAN2-generated faces and cat images (custom dataset created from these sources).

Model(s)

ResNet50

Author countries

USA

← Previous