Wavelet-based GAN Fingerprint Detection using ResNet50

View on arXiv ← Back to list

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

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

AI Summary

This research proposes a wavelet-based method for detecting images generated by StyleGANs by applying Discrete Wavelet Transform (DWT) preprocessing to expose subtle frequency artifacts (GAN fingerprints). By feeding Haar and Daubechies wavelet representations into a ResNet50 classifier, the method achieves significantly higher accuracy compared to standard spatial domain classification. The results demonstrate that frequency-domain analysis effectively uncovers intrinsic artifacts left by the generative process, with the Daubechies-based model achieving 95.1% accuracy.

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 wavelet-based models significantly outperformed the spatial model, which achieved 81.5% accuracy, by reaching 93.8% (Haar) and 95.1% (Daubechies) test accuracy. The Daubechies-based model showed the best performance (95.1% accuracy and 0.97 AUC), confirming that using more complex wavelets captures nuanced frequency patterns indicative of GAN fingerprints. These findings validate that frequency-domain analysis provides a more useful signal for GAN detection than raw spatial features.

Approach

The input images are preprocessed using Discrete Wavelet Transform (DWT) utilizing Haar and Daubechies wavelet filters to convert them into multi-resolution frequency representations. These representations, which emphasize high-frequency artifacts, are then fed into a ResNet50 Convolutional Neural Network for binary classification (real vs. fake). The performance of the wavelet-based models is compared against an identical ResNet50 model trained on raw spatial data.

Datasets

Flickr-Faces-HQ (FFHQ), Cats vs Dogs dataset (Kaggle), and custom StyleGAN2-generated images (Faces and Cats).

Model(s)

ResNet50, combined with Discrete Wavelet Transform (DWT) using Haar and Daubechies wavelets.

Author countries

USA

← Previous