[Paper Review] Unmasking DeepFakes with simple Features

Ricard Durall, Margret Keuper+2 more|arXiv (Cornell University)|Nov 2, 2019

Digital Media Forensic DetectionComputer Science27 references174 citations

TL;DR

The paper detects DeepFakes using a simple frequency-domain feature (1D power spectrum via DFT with azimuthal averaging) plus lightweight classifiers, achieving very high accuracy on high/medium resolutions and robust performance in unsupervised settings.

ABSTRACT

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

Motivation & Objective

• Motivate a lightweight, data-efficient approach to detecting AI-generated fake faces.
• Leverage frequency-domain artifacts to distinguish real from fake images without large labeled datasets.
• Introduce Faces-HQ, a high-resolution real/fake face dataset for evaluation.
• Show robustness across high, medium, and low-resolution data (images and videos).

Proposed method

• Compute the Discrete Fourier Transform on grayscale face images.
• Compute the azimuthal average of the FFT power spectrum to obtain a 1D feature vector (722 features).
• Train simple classifiers (SVM with RBF, Logistic Regression, and K-Means) on the 1D power spectrum features.
• Evaluate under supervised and unsupervised settings across multiple datasets (Faces-HQ, CelebA, FaceForensics++).
• For video data, interpolate the 1D spectrum to a fixed size before classification.

Experimental results

Research questions

• RQ1 Can simple frequency-domain features reveal artifacts of GAN-generated faces across resolutions?
• RQ2 What is the data efficiency and accuracy of lightweight classifiers on 1D power-spectrum features?
• RQ3 How does the method perform on high-, medium-, and low-resolution data (images and videos) compared to prior deep-learning detectors?

Key findings

• High-resolution evaluation on Faces-HQ yields 100% accuracy with as few as 20 annotated samples.
• Medium-resolution CelebA results reach 100% accuracy in supervised learning and 96% in unsupervised settings.
• Low-resolution FaceForensics++ video evaluation achieves 90% accuracy in frame-based detection.
• SVM and Logistic Regression consistently achieve near-perfect performance with sufficient sample sizes; K-Means performs worse but still competitive in some settings.
• Frequency components grouped into sub-sections reveal that certain high-frequency bands drive discrimination (e.g., 100–300 range yielding up to 0.86–1.00 accuracy in some setups).
• The method remains robust across data sources and GAN types, relying on frequency-domain artifacts rather than large labeled training.