[Paper Review] Detecting and Simulating Artifacts in GAN Fake Images

Xu Zhang, Svebor Karaman+1 more|arXiv (Cornell University)|Jul 15, 2019

Digital Media Forensic Detection15 references41 citations

TL;DR

The paper introduces AutoGAN to simulate GAN artifacts and a spectrum-based detector to distinguish GAN-generated images, achieving strong generalization without access to target GAN models.

ABSTRACT

To detect GAN generated images, conventional supervised machine learning algorithms require collection of a number of real and fake images from the targeted GAN model. However, the specific model used by the attacker is often unavailable. To address this, we propose a GAN simulator, AutoGAN, which can simulate the artifacts produced by the common pipeline shared by several popular GAN models. Additionally, we identify a unique artifact caused by the up-sampling component included in the common GAN pipeline. We show theoretically such artifacts are manifested as replications of spectra in the frequency domain and thus propose a classifier model based on the spectrum input, rather than the pixel input. By using the simulated images to train a spectrum based classifier, even without seeing the fake images produced by the targeted GAN model during training, our approach achieves state-of-the-art performances on detecting fake images generated by popular GAN models such as CycleGAN.

Motivation & Objective

• Motivate robust real-vs-GAN fake image detection when the attacker model is unknown.
• Identify artifacts from GAN up-sampling, especially in frequency domain.
• Develop AutoGAN to simulate common GAN pipelines using real images only.
• Propose a spectrum-based classifier to improve generalization across categories.
• Evaluate generalization to unseen GAN models and post-processing effects.

Proposed method

• Analyze GAN generation pipelines to identify up-sampling artifacts in the frequency domain.
• Theoretically show that zero-insertion during up-sampling replicates spectra, motivating spectrum-based inputs.
• Propose a frequency-domain classifier using log-magnitude spectrum of RGB channels as input.
• Introduce AutoGAN, a grey-box GAN simulator that reconstructs images with GAN-like artifacts without fake training data.
• Train classifiers on real and AutoGAN-reconstructed images to capture GAN artifacts.
• Evaluate detection across multiple semantic categories, including leave-one-category-out settings, and test robustness to post-processing and differing up-samplers.

Experimental results

Research questions

• RQ1 Can a spectrum-based input improve GAN fake image detection and generalization across categories compared to pixel-based inputs?
• RQ2 Can a GAN simulator (AutoGAN) train effective detectors without access to actual fake images or targeted GAN models?
• RQ3 How well do spectrum-based detectors generalize to unseen GAN architectures or up-sampling modules?
• RQ4 What is the impact of post-processing (JPEG, resize) on detection and can retraining mitigate it?

Key findings

• Spectrum-based classifiers generalize better across categories than pixel-based counterparts when trained on CycleGAN data.
• AutoGAN enables training without any fake images and still yields strong detection performance on CycleGAN data.
• In leave-one-out evaluations, spectrum-based methods with CycleGAN or AutoGAN training are competitive with state-of-the-art methods.
• Post-processing like JPEG compression or resizing degrades performance unless detectors are retrained with post-processed data.
• Combining up-samplers (Transposed Convolution and Nearest Neighbor) during training improves cross-sampler generalization.
• GauGAN results show that drastic architectural differences limit cross-model generalization, while StarGAN results are more favorable for spectrum-based detectors.