PaperlayerView exact revision on arXiv

Source-linked AI summary

Leveraging Frequency Analysis for Deep Fake Image Recognition

Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, Thorsten Holz

arXiv:2003.08685v3cs.CVeess.IV

TL;DR

GAN-generated images can be difficult for humans to distinguish from real media, motivating better automated detection. This paper analyzes their frequency spectra and finds consistent artifacts that enable accurate, efficient, and robust deep-fake detection.

  • Problem

    GAN outputs can fool humans, making reliable automated detection of deep-fake media an important research problem.

  • Method

    The paper analyzes frequency spectra across GAN architectures, datasets, and resolutions, then tests frequency-based classifiers and upsampling mechanisms.

  • Results

    Frequency-representation classifiers achieve higher accuracy with significantly fewer parameters and greater robustness to common image perturbations.

  • Takeaways & Limitations

    Frequency spectra provide an efficient and accurate basis for separating real from GAN-generated images and identifying their generating GAN.

  • Takeaways & Limitations

    The classifier remains vulnerable to specifically crafted adversarial perturbations that can evade detection.

Abstract

from arXiv · show

Deep neural networks can generate images that are astonishingly realistic, so much so that it is often hard for humans to distinguish them from actual photos. These achievements have been largely made possible by Generative Adversarial Networks (GANs). While deep fake images have been thoroughly investigated in the image domain - a classical approach from the area of image forensics - an analysis in the frequency domain has been missing so far. In this paper, we address this shortcoming and our results reveal that in frequency space, GAN-generated images exhibit severe artifacts that can be easily identified. We perform a comprehensive analysis, showing that these artifacts are consistent across different neural network architectures, data sets, and resolutions. In a further investigation, we demonstrate that these artifacts are caused by upsampling operations found in all current GAN architectures, indicating a structural and fundamental problem in the way images are generated via GANs. Based on this analysis, we demonstrate how the frequency representation can be used to identify deep fake images in an automated way, surpassing state-of-the-art methods.

1. Introduction

GAN-generated media can be difficult for humans to distinguish from real content, motivating automated detection. This paper addresses a gap in frequency-domain analysis by identifying common GAN artifacts, tracing them to upsampling, and using frequency representations for efficient deep-fake detection.

  • Motivation: GAN outputs can be astonishingly effective at fooling humans, with reported performance on a real-versus-generated-image task peaking around 75% accuracy.The task was demonstrated through a website showing one real and one generated image.
  • Motivation: Automated detection is crucial because fake news and Internet information campaigns may influence democratic processes, while suspected deep fakes can affect public communication.The introduction illustrates this concern with a video released during the illness-related absence of Gabon’s president Ali Bongo that was immediately labeled a deep fake.
  • Research gap: Previous detection research used complex image-domain convolutional networks or hand-crafted frequency-domain features, whereas this work analyzes frequency spectra across GAN architectures and data sets.The supplied passage contrasts these approaches with the paper’s comprehensive frequency-spectrum analysis.
  • Detection approach: Frequency representations enable efficient real-versus-fake separation and identification of the specific GAN that generated a sample.The real-versus-fake data are linearly separable in frequency space, while the proposed classifiers achieve higher accuracy with significantly fewer parameters and greater robustness to common image perturbations than state-of-the-art approaches.
  • Key findings: Severe frequency artifacts occur consistently across GAN architectures, data sets, and resolutions, and experiments attribute them to upsampling operations used in current GAN architectures.The paper characterizes this as a structural problem in mapping low-dimensional latent spaces to higher-dimensional inputs.

2. Related Work

This section situates the approach within GAN-based image synthesis and image forensics. It reviews prior image- and frequency-domain methods for detecting GAN-generated images, including concurrent findings on systematic generation flaws.

  • Generative Adversarial Networks: GANs model training-data distributions through a game between a generator and a discriminator.The generator and discriminator are usually neural networks.
  • Image Synthesis: The paper focuses on GAN-generated images, whose synthesis advanced substantially after the introduction of convolutional neural networks.Subsequent work introduced techniques including spectral normalization, progressive growing, style mixing, and very large models.
  • Image Forensics: Traditional image forensics detects tampered media using natural image statistics, while steganalysis uses high-frequency residuals to identify manipulations.CNN-based methods later extended these approaches by learning more complex features.
  • Image Forensics: Prior systems identify GAN-generated images with steganalysis, CNNs, image-domain statistics, or handcrafted frequency-domain features.Marra et al. compare steganalysis and CNN-based methods, while other approaches operate specifically in the image or frequency domain.
  • Related Detection Work: Concurrently, Wang et al. and Durall et al. reported related observations about detectable artifacts or systematic flaws in CNN-generated images.Wang et al. found that a classifier trained on images from one CNN generator could generalize to unseen architectures, data sets, and training methods when carefully augmented.

3. Frequency Artifacts

DCT analysis reveals distinctive frequency-domain artifacts in GAN-generated images, including strong high-frequency components and grid-like patterns. These artifacts appear across GAN architectures and are linked to upsampling operations, becoming less severe with more elaborate techniques and larger kernels.

  • DCT analysis: The study transforms images into the frequency domain with the discrete cosine transform (DCT), whose coefficient magnitudes indicate contributions from corresponding spatial frequencies.DCT spectra are visualized as heatmaps, with low frequencies in the upper-left and high frequencies in the lower-right.
  • Observed artifacts: StyleGAN images contain strong high-frequency components, generally higher spectral magnitudes, and grid-like patterns compared with natural-image spectra.The artifacts are visible throughout the spectrum, including regions associated with the highest frequencies along the x and y directions.
  • Cross-architecture analysis: Mean spectra from BigGAN, ProGAN, StyleGAN, and SN-DCGAN frequently show grid-like patterns, indicating a structural problem in GAN image generation.StyleGAN more closely approximates natural-image spectra than the other examined GANs but still has elevated coefficients along the upper and left sides.
  • Artifact origin: More elaborate upsampling techniques and larger kernels produce smoother spectral images and less severe artifacts.These results support the hypothesis that upsampling operations cause the observed frequency-domain artifacts.

4. Frequency-Based Deep-Fake Recognition

Frequency-domain representations make GAN-generated images readily distinguishable from real images, achieving perfect separation in the StyleGAN experiment and improving classifiers across architectures and datasets. They also require less data, converge faster, and are generally more robust to perturbations, although noise and adversarially crafted perturbations remain limitations.

  • Frequency-domain separation: DCT representations perfectly separate real FFHQ images from StyleGAN images, reaching 100 % test accuracy with a linear classifier.Raw-pixel classification requires non-linear models, whereas frequency-domain artifacts make linear separation possible.
  • Upsampling artifacts: Heavier anti-aliasing reduces classifier accuracy, while lower-resolution samples require fewer upsampling operations and exhibit fewer artifacts.Despite reduced visible artifacts, binomial upsampling still leaves a grid-like frequency structure used for classification.
  • Architecture classification: Frequency-domain training improves every tested classifier, with kNN gaining roughly 42 % and Eigenfaces improving by 47.24 % and 30.81 %.The shallow CNN also improves by +0.69 % and 1.27 % on the two datasets.
  • Architecture classification: CNN error rates fall from 1.05 % and 2.20 % to 0.36 % and 0.93 %, corresponding to reductions of 66 % and 50 %.For Yu et al.’s CNN, error rates similarly drop from 1.76 % and 0.3 % to 0.39 % and 0.09 %.
  • Data efficiency and training: With only 20 % of the original data, frequency-domain classifiers retain high accuracy, whereas raw-pixel classifiers deteriorate significantly.DCT variants also converge significantly faster on both CelebA and LSUN for both classifiers.
  • Robustness and limitations: DCT classifiers are more robust to all tested perturbations except noise, but specifically crafted adversarial perturbations can evade detection.The adversarial-evasion limitation was demonstrated by subsequent work.

5. Discussion and Conclusion

The study finds that GAN-generated images exhibit frequency-spectrum artifacts consistent across architectures, datasets, and resolutions, and links them to upsampling operations. It also discusses mitigation strategies, noting that removing upsampling is impractical and DCT-based methods were unsuccessful.

  • Main findings: Frequency-spectrum artifacts occur consistently across GAN architectures, datasets, and resolutions, and experiments confirm that upsampling operations cause them.The frequency representation can therefore support efficient and accurate deep-fake identification.
  • Limitations and mitigation: Removing upsampling could mitigate artifacts, but would eliminate compact latent spaces and require infeasible full-resolution training.A StyleGAN instance already requires 26.2M parameters, making this approach impractical for the foreseeable future.
  • Limitations and mitigation: DCT-based discriminators and generator penalties failed to improve results, either having no effect when weakly weighted or causing training collapse when dominant.The authors leave further exploration of these methods open.

A. FFHQ

For FFHQ, StyleGAN images show frequency-spectrum artifacts and diverge most from real images at higher frequencies. A LASSO classifier exploits these high-frequency differences, which can be associated with upsampling operations.

  • A. FFHQ: StyleGAN images contain additional frequency-spectrum artifacts scattered away from the grid when compared with real FFHQ images.The real and generated spectra were compared using averages over 10,000 images.
  • A. FFHQ: Generated FFHQ images diverge most from the real spectrum at higher frequencies, where real images contain little energy.These high frequencies can also be attributed to upsampling operations.
  • A. FFHQ: A LASSO regression trained on FFHQ focuses primarily on higher frequencies, especially the top-right and lower-left corners of the spectrum.The classifier’s frequency focus was visualized by mapping its weight vector back to the corresponding frequencies.

B. Kaggle

The section analyzes mean DCT frequency spectra for Stanford dog images and images generated by several GAN instances. The estimate E[D(I)] is computed by averaging over 10,000 images, with spectra shown separately by color channel.

  • B. Kaggle: Mean DCT spectra are compared between the Stanford dog data set and images generated by BigGAN, ProGAN, StyleGAN, and SN-DCGAN.The GANs were trained on the Stanford dog data set.
  • B. Kaggle: 10,000 images are used to estimate E[D(I)] by averaging across the samples.
  • B. Kaggle: Figure 10 presents the frequency spectra of sample sets from different GAN types, split into color channels.

C. Upsampling

The mean DCT spectra of images produced with different upsampling techniques reveal frequency artifacts, including artifacts that persist with binomial upsampling. The analysis estimates spectra over 10,000 sampled images and compares them with the training-image mean spectrum.

  • Upsampling: Mean DCT spectra are computed for images generated using different upsampling techniques and compared with the training-image mean spectrum.The comparison includes the absolute difference from the training-image mean spectrum.
  • Upsampling: 10,000 images are averaged to estimate E[D(I)] for each corresponding network or training-data distribution.The estimate is formed separately for images sampled from the network and from the training data.
  • Upsampling: Binomial upsampling reduces the grid pattern but still leaves artifacts scattered throughout the frequency spectrum.This observation is made alongside spectra from other upsampling techniques.

D. Network Architecture

The CNN uses TensorFlow’s standard Adam configuration and a batch size of 1024. Experiments found that changing optimizer settings or reducing batch size did not substantially affect training.

  • Training configuration: The CNN is trained with Adam using an initial learning rate of 0.001, β1 = 0.9, β2 = 0.999, and ϵ = 1−7.These are described as the standard parameters for a TensorFlow implementation.
  • Training configuration: Experiments with different optimizer settings did not substantially influence training, so the standard configuration was retained.The passage does not specify the alternative settings tested.
  • Training configuration: The network is trained with a batch size of 1024 after lower batch sizes showed no influence on training.The largest batch size was selected.
Loading 2003.08685v3…