Source-linked AI summary
On the generalization of GAN image forensics
Xinsheng Xuan, Bo Peng, Wei Wang, Jing Dong
TL;DR
The paper addresses limited generalization of GAN-image forensic detectors to unseen generators, an important issue as GAN types proliferate. It preprocesses real and fake training images with blur or noise so a CNN learns more intrinsic features, then evaluates the trained model on original images. Experiments show improved generalization on unseen fake-image types, although the gain is modest.
Problem
Existing detectors often target one generated-image type, leaving generalization to unseen GAN-generated images insufficiently addressed.
Method
The method applies image preprocessing such as Gaussian blur or noise to real and fake training images, forcing a forensic CNN toward intrinsic features rather than unstable generator-specific cues.
Results
The method improves generalization on unseen generated-image datasets, including about 10 percentage points of TNR improvement in two comparisons, while same-type performance remains nearly constant after preprocessing.
Takeaways & Limitations
Preprocessing can improve a forensic CNN’s detection generalization without using target-domain fake images during training.
Takeaways & Limitations
Generalization remains difficult because different generator models can produce substantially different image distributions, and the reported performance increment is not large.
Abstract
from arXiv · showhide
Recently the GAN generated face images are more and more realistic with high-quality, even hard for human eyes to detect. On the other hand, the forensics community keeps on developing methods to detect these generated fake images and try to guarantee the credibility of visual contents. Although researchers have developed some methods to detect generated images, few of them explore the important problem of generalization ability of forensics model. As new types of GANs are emerging fast, the generalization ability of forensics models to detect new types of GAN images is absolutely an essential research topic. In this paper, we explore this problem and propose to use preprocessed images to train a forensic CNN model. By applying similar image level preprocessing to both real and fake training images, the forensics model is forced to learn more intrinsic features to classify the generated and real face images. Our experimental results also prove the effectiveness of the proposed method.
1 Introduction
GAN-generated faces can be difficult to distinguish from real images, creating forensic and security risks. Existing detectors often fail to address generalization to unseen GAN types, motivating preprocessing-based CNN training.
- Motivation: GANs can generate high-quality face images that human observers may not distinguish from real images.Such images can support DeepFake manipulation, fake news, and attacks on face-recognition security.
- Motivation: Generated face images create forensic risks because fake news and deceptive face-recognition inputs can harm communities and system security.
- Research Gap: Existing detection methods largely target one generated-image type, leaving performance on unseen types insufficiently addressed.
- Proposed Direction: The paper trains a forensic CNN with Gaussian blur or noise applied to real and fake images to encourage learning intrinsic rather than generator-specific features.The preprocessing suppresses unstable low-level high-frequency cues and aims to improve generalization.
2 Related Work
Prior work detects AI-generated images and videos, but most methods evaluate only on the same generation type used for training. The paper instead targets generalization without requiring fake examples from an unknown target generator.
- Existing Methods: Deep networks and ensemble or color-statistical methods have been proposed for detecting generated videos and GAN face images.
- Generalization Gap: Many prior forensic methods train and test on the same generated-image type, leaving generalization to new GAN models unknown.
- Comparison: ForensicTransfer adapts to new domains and can use a handful of target-domain fake examples, unlike this paper’s setting without target-domain fake images.
3 Proposed Method
The proposed method preprocesses training images to suppress unstable low-level artifacts and force the forensic CNN toward more intrinsic features. Preprocessing is used during training but omitted at testing, where original images are classified.
- Training Pipeline: The training pipeline applies smoothing filtering or noise before the CNN to reduce unstable high-frequency cues and improve feature generalization.This intentionally contrasts with forensic networks designed to enhance high-frequency pixel noise.
- Testing Pipeline: At testing, the preprocessing step is removed and the trained network directly receives original images.Training includes unchanged cases through blur kernel 1 and zero-variance noise.
- Preprocessing: Gaussian blur and Gaussian noise alter low-level pixel statistics, with randomly varied preprocessing strength increasing training-sample diversity.Blur kernel sizes are randomly selected from 1, 3, 5, and 7 for each training batch.
- Network Architecture: The approach uses a simple DCGAN discriminator as a binary CNN classifier rather than designing a complex architecture.It uses four stride-2 convolutional layers with 4x4 kernels, batch normalization except in the first layer, Leaky ReLU activations, binary cross-entropy, and Adam.
4 Experiments
Experiments train on CelebA-HQ real faces and PGGAN fakes, then test on matched and unseen GAN-generated datasets. Preprocessing preserves in-domain performance and improves generalization to unseen fake-image types, though the task remains difficult.
- 4.1 Experimental Setups: The CNN is trained on CelebA-HQ real images and PGGAN fake images, while DCGAN and WGAN-GP datasets test generalization to unseen generators.All images are resized to 128x128 for the CNN; DCGAN and WGAN-GP are used only for testing.
- 4.2 Improvement of Model Generalization: Accuracy, true positive rate, and true negative rate remain almost constant after Gaussian blur or Gaussian noise preprocessing on the matched test domain.The baseline M is compared with MGB and MGN preprocessing models.
- 4.2 Improvement of Model Generalization: More than 95% ACC, TPR, and TNR are achieved when testing on the same generated-image type used for training.Performance is substantially lower on the unseen WGAN-GP and DCGAN test datasets.
- 4.2 Improvement of Model Generalization: Around 10% TNR improvement and higher overall ACC are observed on unseen WGAN-GP images after preprocessing.A similar approximately 10% TNR improvement is reported for the unseen DCGAN dataset.
- 4.2 Improvement of Model Generalization: Generalization remains challenging because different generator models can produce fake-image distributions that differ substantially.The paper illustrates real and fake distributions in a simplified two-dimensional feature space.
5 Conclusion
The paper investigates generalization in GAN-image detection and proposes preprocessing during training to encourage more intrinsic, generalizable features. Experiments show preliminary improvement, but generalization remains difficult because fake-image distributions vary across generators.
- 5 Conclusion: The paper studies how to improve a forensic CNN’s ability to detect generated images from unseen GAN models.Existing detection models often focus on the generator type used during training and testing.
- 5 Conclusion: The proposed method preprocesses training images before CNN training to force the discriminator to learn more intrinsic and generalizable features.The paper presents this as a strategy distinct from existing approaches.
- 5 Conclusion: Experiments show that the approach improves generalization, although the performance increment is not large.The authors describe the results as preliminary because of the inherent difficulty of the problem.
- 5 Conclusion: Fake-image distributions may differ substantially across generator models, making generalization to future unknown generators difficult.The paper positions this distribution variation as a central challenge for future work.