PaperlayerView exact revision on arXiv

Source-linked AI summary

Unmasking DeepFakes with simple Features

Ricard Durall, Margret Keuper, Franz-Josef Pfreundt, Janis Keuper

arXiv:1911.00686v3cs.LGcs.CVstat.ML

TL;DR

Realistic AI-generated face images are increasingly difficult to distinguish from authentic content, creating a need for effective detection methods. This paper uses frequency-domain analysis with a simple classifier and reports 100% accuracy on high- and medium-resolution images and 91% on low-resolution videos.

  • Problem

    The growing realism and accessibility of deep generative models has created a need to detect artificial face images.

    Source: §I · p. 1

  • Method

    The method analyzes high-frequency image components in the frequency domain and feeds the resulting features to a simple supervised or unsupervised classifier.

    Source: §I · pp. 1–2

  • Results

    100% accuracy was achieved on high- and medium-resolution deepfake images, while low-resolution manipulated videos reached 91% accuracy.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    The approach detects high- and medium-resolution deepfakes accurately across datasets, while low-resolution content remains harder to identify.

    Source: §V · p. 7

Abstract

from arXiv · show

Deep generative models have recently achieved impressive results for many real-world applications, successfully generating high-resolution and diverse samples from complex datasets. Due to this improvement, fake digital contents have proliferated growing concern and spreading distrust in image content, leading to an urgent need for automated ways to detect these AI-generated fake images. Despite the fact that many face editing algorithms seem to produce realistic human faces, upon closer examination, they do exhibit artifacts in certain domains which are often hidden to the naked eye. In this work, we present a simple way to detect such fake face images - so-called DeepFakes. Our method is based on a classical frequency domain analysis followed by basic classifier. Compared to previous systems, which need to be fed with large amounts of labeled data, our approach showed very good results using only a few annotated training samples and even achieved good accuracies in fully unsupervised scenarios. For the evaluation on high resolution face images, we combined several public datasets of real and fake faces into a new benchmark: Faces-HQ. Given such high-resolution images, our approach reaches a perfect classification accuracy of 100% when it is trained on as little as 20 annotated samples. In a second experiment, in the evaluation of the medium-resolution images of the CelebA dataset, our method achieves 100% accuracy supervised and 96% in an unsupervised setting. Finally, evaluating a low-resolution video sequences of the FaceForensics++ dataset, our method achieves 91% accuracy detecting manipulated videos. Source Code: https://github.com/cc-hpc-itwm/DeepFakeDetection

I. INTRODUCTION

The paper targets increasingly accessible, high-quality AI-generated face manipulation, introducing a frequency-domain detection pipeline with simple supervised or unsupervised classification and limited data requirements. It also contributes the Faces-HQ dataset and reports strong accuracy across image and video benchmarks.

Motivation and problem

  • Deep generative models have enabled realistic, high-quality fake images to spread rapidly, while existing detectors often require large amounts of labeled training data.Variational Autoencoders and Generative Adversarial Networks are identified as prominent generative approaches [16],.Source: §I · p. 1

Detection method

  • The method detects artificial faces through classical frequency-domain analysis followed by a simple supervised or unsupervised classifier.The approach analyzes differing high-frequency behavior and does not require vast quantities of labeled data.Source: §I · pp. 1–2

Contribution

  • The paper introduces a novel artificial-face classification pipeline based on frequency-domain analysis.The pipeline transforms images into frequency-based features before classification.Source: §I · pp. 1–2
  • Faces-HQ is provided as a public dataset of high-quality real and fake faces assembled from multiple public databases.It complements the CelebA and FaceForensics++ datasets [25] for evaluation.Source: §I · p. 2

Evaluation results

  • 100% accuracy was achieved on high- and medium-resolution Faces-HQ and CelebA images, while FaceForensics++ low-resolution videos reached 91%.The reported experiments cover high- and medium-resolution images and low-resolution video sequences.Source: §I · p. 2

II. RELATED WORK

The paper reviews image-forensics methods for detecting GAN-generated images and DeepFakes, spanning handcrafted features, neural networks, artifact-based cues, and color-space analysis. It also highlights that differentiable deep-learning detectors can be learned and countered by the GAN generator.

Traditional and neural forensics

  • Traditional forensics targets local noise, patterns, illumination, or steganalysis features, whereas CNN methods learn related cues implicitly,,,, [8], [28].The field shifted substantially toward neural-network techniques after the deep-learning breakthrough.Source: §II · p. 2

GAN-generated media

  • GANs, introduced by Goodfellow et al. in 2014, substantially advanced artificial-face generation [5] and prompted growing efforts to detect resulting DeepFake images and videos.These manipulations became established online as generative image quality improved.Source: §II · p. 2

Artifact-based detection

  • Existing detectors exploit artifacts such as missing eye blinks [18], unnatural head poses [26], or disparities in color-space features between generated and real images [17], [22].The eye-blink cue can be circumvented by adding closed-eye images to training.Source: §II · p. 2

Deep-learning detectors

  • CNN-based detectors [2],, [27], contrastive deep forgery discriminators [13], and CNN-RNN temporal models show promise but may be countered by generators trained against differentiable forensic detectors.Embedding a detector in the GAN discriminator can let the generator learn a countermeasure.Source: §II · p. 2

III. METHOD · A. Frequency Domain Analysis · 1) Discrete Fourier Transform:

The method analyzes images in the Fourier domain, where spectral decomposition represents how signal energy is distributed across frequencies. It then applies the Discrete Fourier Transform to obtain frequency-specific amplitude and phase information.

III. METHOD

  • The method section introduces the processing pipeline overview shown in Figure 2.Source: §III · p. 2

A. Frequency Domain Analysis

1) Discrete Fourier Transform:

  • The Discrete Fourier Transform decomposes a discrete signal into sinusoidal components from zero frequency, representing the image mean, to the maximum frequency allowed by spatial resolution.For 2-dimensional data of size M × N, the DFT is the discrete counterpart of the continuous Fourier Transform for uniformly sampled signals.Source: 1) Discrete Fourier Transform · p. 3
  • The frequency-domain representation records amplitude and phase at each frequency, with Fig. 3 illustrating the power and phase spectra and amplitude equal to the square root of power.Source: 1) Discrete Fourier Transform · p. 3, Figure 3 · p. 3

2) Azimuthal Average: · B. Classifier Algorithms

The method compresses the 2D FFT power spectrum into a robust 1D feature vector through azimuthal averaging, then frames classification as learning mappings from attribute vectors to discrete classes. Input images are converted to grayscale before the FFT.

2) Azimuthal Average:

  • Azimuthal averaging produces a robust 1D representation of the 2D FFT power spectrum by radially averaging frequency components.The radial averages form the compressed power-spectrum vector shown in Fig. 4.Source: 2) Azimuthal Average · p. 3, Figure 4 · p. 3
  • The frequency compression reduces feature dimensionality while preserving relevant information and improving representation robustness.Similar frequency components are gathered and averaged into a feature vector.Source: 2) Azimuthal Average · p. 3
  • The Fourier-transformed image remains two-dimensional before azimuthal averaging because the transformation preserves dimensionality.Azimuthal averaging is therefore applied to convert the 2D spectral information into one dimension.Source: 2) Azimuthal Average · p. 3

B. Classifier Algorithms

1) Logistic Regression: · 2) Support Vector Machines:

The paper introduces Logistic Regression as a simple probabilistic binary classifier and Support Vector Machines as margin-based classifiers for nonlinear data. Both methods are defined through optimization procedures that determine their separating model parameters.

1) Logistic Regression:

  • Logistic Regression uses a logistic function to model a binary dependent variable and outputs an estimated probability for each input.The hypothesis output indicates confidence in the predicted value.Source: 1) Logistic Regression · p. 3
  • Maximum likelihood estimation determines the Logistic Regression coefficient w to fit binary-outcome probabilities.Training stops when convergence is reached or the maximum iteration limit is met.Source: 1) Logistic Regression · p. 3

2) Support Vector Machines:

  • Support Vector Machines (SVMs), are widely used learning algorithms for nonlinear data classification.Their formulation learns a classifier from training instance-label pairs.Source: 2) Support Vector Machines · p. 3
  • SVM training seeks an optimal separating hyperplane that maximizes the margin between different classes.The classifier is parameterized by w and b.Source: 2) Support Vector Machines · pp. 3–4
  • The SVM optimization includes slack variable ξ and penalty parameter C > 0 for the error term.These parameters appear alongside the classifier parameters w and b in the optimization problem.Source: 2) Support Vector Machines · p. 3
  • SVMs map training vectors x_i into a higher-dimensional space using φ, where they find a linear hyperplane with maximal margin.The feature-space transformation enables the stated linear separation objective.Source: 2) Support Vector Machines · pp. 3–4

3) K-Means Clustering: · IV. EXPERIMENTS

The paper evaluates K-means as an unsupervised alternative for grouping feature representations and tests the approach across Faces-HQ, CelebA [19], and FaceForensics++ [25]. K-means assigns nearby features to clusters and updates centroids, with user-specified K enabling classification by class-based grouping.

3) K-Means Clustering:

  • The experiments assess detection without labeled training examples by using unsupervised clustering to group similar feature representations.This complements supervised SVM and LR classifiers that require labeled examples.Source: 3) K-Means Clustering · p. 4
  • K-means relies on the assumption that nearby points in feature space have similar qualities and can therefore be clustered together.Source: 3) K-Means Clustering · p. 4
  • The K-means objective function is defined for clustering the feature representations.Source: 3) K-Means Clustering · p. 4
  • K-means iteratively assigns features to the nearest initial centroid and reestimates centroids from those assignments.This provides a heuristic approximation to the clustering solution.Source: 3) K-Means Clustering · p. 4
  • Because the user specifies the number of clusters, K-means can support classification by dividing data into K clusters with K equal to or greater than the number of classes.Source: 3) K-Means Clustering · p. 4

IV. EXPERIMENTS

  • The experiments introduce the high-resolution Faces-HQ dataset and evaluate the approach on medium-resolution CelebA [19] images and low-resolution FaceForensics++ [25] video sequences.The section presents training settings and detailed results for Faces-HQ before evaluating the other datasets.Source: §IV · p. 4

A. Faces-HQ · 1) Data set: · 2) Training Setting:

The authors introduce Faces-HQ2, a 40K-image high-resolution benchmark balancing real and fake faces from four established sources. Their training pipeline converts images into compact 1D power-spectrum features, then trains and evaluates a binary classifier using an 80/20 split.

1) Data set:

  • Faces-HQ2 contains 40K high-quality face images, evenly split between real and fake samples.The dataset addresses the lack of public high-resolution data with annotated real and fake faces.Source: 1) Data set · p. 4
  • The dataset combines images from CelebA-HQ [14], Flickr-Faces-HQ [15], the 100K Faces project, and thispersondoesnotexist.com.Source: 1) Data set · p. 4
  • Faces-HQ2 occupies 19GB of storage.Source: 1) Data set · p. 4

2) Training Setting:

  • Each 1024x1024x3 color image is transformed from the spatial domain into 722 one-dimensional frequency-domain features.The transformation uses a Discrete Fourier Transform followed by azimuthal averaging.Source: 2) Training Setting · p. 4
  • The pipeline retains only the power spectrum, which the authors report contains sufficient information for classification.The Fourier transformation can be optimized with the Fast Fourier Transform.Source: 2) Training Setting · p. 4
  • A t-SNE visualization reveals clear clustering of fake and real samples in the power-spectrum feature space.The visualization uses a random Faces-HQ subset with perplexity 4 and 4000 iterations.Source: 2) Training Setting · p. 4, Figure 5 · p. 4
  • The transformed data are split into 80% training and 20% testing, after which a binary classifier is trained to distinguish real from fake faces.Accuracy is evaluated on the held-out testing set.Source: 2) Training Setting · p. 4

3) Method 1D Power Spectrum:

The 1D Power Spectrum reveals class-specific patterns and distinct high-frequency behavior between real and fake faces, enabling straightforward classification. Experiments evaluate this representation with multiple classifiers and frequency subsets under varying data sizes.

3) Method 1D Power Spectrum:

  • Real and fake images show noticeably different high-frequency spectra, with repetitive within-class patterns that enable easy classification.Real faces lack flat high-frequency regions, whereas fake faces exhibit them.Source: 3) Method 1D Power Spectrum · p. 4, Figure 6 · p. 5
  • The method computes basic statistics over 4,000 images and evaluates SVM and logistic regression classifiers using 80% training and 20% testing across sample sizes of 4,000, 1,000, 100, and 20.The samples are equally distributed across the sub-datasets.Source: 3) Method 1D Power Spectrum · p. 5
  • The 722 Power Spectrum features are grouped into 28 frequency sections to measure the relevance of different frequency components with SVM, logistic regression, and k-means.Tables III–V report accuracy according to the start and end frequencies of each section.Source: 3) Method 1D Power Spectrum · p. 5, Table III · p. 5, Table IV · p. 6
  • 0.86 accuracy is obtained by a frequency chunk containing frequencies from 100 to 300.The chunk-based analysis tests how selected frequency ranges affect detection performance.Source: 3) Method 1D Power Spectrum · p. 6

B. CelebA · 1) Data set: · 2) Training Setting:

The CelebA experiments use 202,599 medium-resolution celebrity face images with 40 facial-attribute variations. Real images are paired with DCGAN-generated fakes, split for training and testing, then resized for classification.

B. CelebA

  • 202,599 celebrity face images comprise the CelebA dataset, which includes 40 facial-attribute variations.Source: 1) Data set · p. 6

1) Data set:

  • CelebA images have dimensions 178x218x3, representing medium resolution in this experimental context.Source: 1) Data set · p. 6

2) Training Setting:

3) Results: · C. FaceForensics++ · 2) Training Setting:

On FaceForensics++ [25], the method uses frame extraction, inner-face cropping, fixed-size normalized power-spectrum features, and achieves perfect supervised accuracy with strong unsupervised clustering results.

3) Results:

  • Perfect classification accuracy is achieved in supervised training, with very good results also obtained through unsupervised clustering.The result is reported for SVM, logistic regression, and k-means evaluation.Source: 3) Results · p. 6, Table VI · p. 6

C. FaceForensics++

  • FaceForensics++ [25] contains modified video sequences generated by different automated face-manipulation methods.The dataset also hosts the DeepFakeDetection dataset.Source: C. FaceForensics++ · p. 6
  • Examples show that the manipulations affect only the inner face, contrasting one real face with two deepfake faces.The figure provides cropped samples and corresponding statistics from the DeepFakeDetection dataset.Source: C. FaceForensics++ · p. 6, Figure 7 · p. 6
  • The DeepFakeDetection dataset contains 363 original sequences from 28 paid actors in 16 scenes, over 3000 manipulated videos, and corresponding binary masks.All videos contain trackable, mostly frontal faces without occlusions, enabling automated tampering methods to generate realistic forgeries.Source: C. FaceForensics++ · p. 6

2) Training Setting:

  • For video inputs, the pipeline first extracts frames and crops the inner faces before applying the existing processing stages.Different video scenes produce cropped faces with varying sizes.Source: 2) Training Setting · p. 6
  • A new block interpolates each 1D Power Spectrum to 300 features and normalizes it by the 0th frequency component before classification.The size-independent preprocessing remains unchanged, while the added block supplies the fixed feature dimension required by the classifiers.Source: 2) Training Setting · p. 6

3) Method 1D Power Spectrum: · V. DISCUSSION AND CONCLUSION

The method detects AI-generated face images through high-frequency component analysis, achieving perfect accuracy on high- and medium-resolution datasets and 91% accuracy on low-resolution benchmark content. The discussion attributes lower low-resolution performance to the smaller available frequency spectrum.

3) Method 1D Power Spectrum:

  • The 1D power-spectrum experiments examine deepfake images and classifier accuracy under different data settings.The supplied figure and tables cover 1D power-spectrum statistics and test accuracy using SVM and logistic regression classifiers, including per-video accuracy.Source: 3) Method 1D Power Spectrum · p. 6, Table VII · p. 7, Figure 8 · p. 7

V. DISCUSSION AND CONCLUSION

  • 100% accuracy was achieved for high- and medium-resolution deepfake images across two datasets containing data from various GANs.The experiments were intended to demonstrate robustness independently of the source image.Source: §V · p. 7
  • 91% accuracy was achieved when identifying low-resolution fakes in a popular benchmark.The method still detected low-resolution fakes despite the reduced available frequency spectrum.Source: §V · p. 7
  • Low-resolution deepfake detection is harder because the available frequency spectrum is much smaller.This limitation materially affects the method’s performance on low-resolution content.Source: §V · p. 7
  • The proposed pipeline exposes AI-generated fake face images using high-frequency component analysis.The paper describes and evaluates the efficacy of this new detection method.Source: §V · p. 7
  • The experiments evaluate whether the detection pipeline remains robust independently of the source image.The paper reports extensive experiments supporting this robustness claim.Source: §V · p. 7

Related papers

Loading 1911.00686v3…