Source-linked AI summary
Incremental learning for the detection and classification of GAN-generated images
Francesco Marra, Cristiano Saltori, Giulia Boato, Luisa Verdoliva
TL;DR
Hyper-realistic GAN images make it difficult to distinguish generated from real content, while new architectures continually challenge detectors. This paper adapts incremental learning to jointly detect and classify GAN-generated images, and experiments show robust detection and classification as new GANs are introduced.
Problem
Existing CNN detectors perform well on familiar images but often fail on images generated by different architectures, creating a need for detectors that adapt as GANs evolve.
Method
The paper adapts iCaRL-based incremental learning with exemplar memory and multi-task objectives for simultaneous GAN-image detection and architecture classification.
Results
Detection accuracy remains over 90% as architectures increase, while classification accuracy exceeds 90% for almost all architectures and 80% for ProGAN and Glow.
Takeaways & Limitations
The method detects newly introduced GANs without reducing performance on previously encountered GANs and also supports identifying the generating architecture.
Abstract
from arXiv · showhide
Current developments in computer vision and deep learning allow to automatically generate hyper-realistic images, hardly distinguishable from real ones. In particular, human face generation achieved a stunning level of realism, opening new opportunities for the creative industry but, at the same time, new scary scenarios where such content can be maliciously misused. Therefore, it is essential to develop innovative methodologies to automatically tell apart real from computer generated multimedia, possibly able to follow the evolution and continuous improvement of data in terms of quality and realism. In the last few years, several deep learning-based solutions have been proposed for this problem, mostly based on Convolutional Neural Networks (CNNs). Although results are good in controlled conditions, it is not clear how such proposals can adapt to real-world scenarios, where learning needs to continuously evolve as new types of generated data appear. In this work, we tackle this problem by proposing an approach based on incremental learning for the detection and classification of GAN-generated images. Experiments on a dataset comprising images generated by several GAN-based architectures show that the proposed method is able to correctly perform discrimination when new GANs are presented to the network
I. INTRODUCTION
Hyper-realistic GAN images increasingly challenge the distinction between real and generated content. Existing detectors must adapt as new GAN architectures and increasingly realistic data appear, motivating incremental detection and classification without degrading performance on earlier GANs.
- GANs can generate photorealistic images and manipulate attributes such as hair color or facial expression, especially convincingly for faces.
- Generated content can deceive human inspection while its broad availability raises concerns about content trustworthiness and misuse.
- Effective detectors should discriminate generated from real content while following improvements in generation quality and realism.
- Prior methods exploit facial artifacts, color information, or fingerprints associated with GAN architectures and specific training instances.
- CNN-based detectors perform well, but new GAN architectures require repeated retraining or fine-tuning as training data continually grows.
- The proposed multi-task incremental method targets detection and classification of new GAN images without worsening performance on previous ones.
II. INCREMENTAL CLASSIFIER USING ICARL
iCaRL incrementally updates a CNN-based classifier using exemplar images and a fixed memory budget to retain knowledge of earlier classes. Its loss combines learning from new data with distillation on exemplars, while exemplar selection and reduction preserve representative old-class information.
- iCaRL adapts classifiers to appearing classes without retraining on the whole dataset by retaining exemplar images from existing classes.
- The memory budget M = |P| remains fixed, limiting the number of stored exemplars as new classes are added.
- During initialization, a CNN produces unit-norm feature vectors, and M/s images per class form exemplar sets whose averages define class templates.
- When new classes arrive, iCaRL updates the CNN with new data and existing exemplars, creates new exemplar sets, and reduces old sets to preserve M.
- The objective balances cross-entropy classification on new data with a distillation term on exemplars through weight γ.
- The paper extends iCaRL to GAN detection and classification through two multi-task variants, including a separate detector or an added binary loss.
- Exemplars are selected near class-average feature vectors, while old exemplars farthest from their class template are discarded.
III. PROPOSED METHOD
The method adapts iCaRL to jointly detect GAN-generated images and classify their generating architectures as new GAN classes appear. It proposes two multi-task variants that preserve prior knowledge through exemplar memory while adding detection-specific objectives.
- Incremental formulation: The method adapts iCaRL for GAN forensics, aiming to detect and classify new generated images without degrading performance on previously learned architectures.The adaptation targets incremental updates while retaining old-class recognition.
- Incremental update: A fixed exemplar memory supports incremental updates while limiting the stored data used to preserve earlier classes.The described update considers new training sets and exemplars while maintaining the memory budget.
- Detection rationale: The classification-based formulation is intended to address both existing and upcoming GAN architectures without relying on a direct universal GAN detector.The authors state that no reliable GAN “smoking gun” had been identified to enable direct detection.
- Problem formulation: The approach treats each GAN architecture as a separate class and uses incremental learning to update classification and detection when new architectures appear.Each architecture is associated with GAN-generated and real classes, forming a forensic multi-task problem.
- MT-MC: MT-MC combines an incremental classifier with a separate binary detector trained through an additional detection loss.The detector operates in parallel with the classifier.
- MT-SC: MT-SC uses one classifier for both tasks and adds a binary cross-entropy term to encourage GAN-versus-real detection.The additional loss is computed from classifier activations for GAN and real classes and jointly back-propagated.
IV. EXPERIMENTAL RESULTS
The experiments evaluate the proposed method on a publicly available GAN dataset, reporting separate detection and classification results against state-of-the-art approaches.
- Evaluation scope: The proposed method is validated on a publicly available GAN dataset through experiments covering GAN detection and classification.The evaluation compares both tasks with state-of-the-art approaches.
A. GAN-image dataset
The GAN dataset contains real and generated images associated with five established generative architectures, with fixed training, validation, and test splits.
- Dataset composition: The dataset uses CycleGAN, ProGAN at 256×256 and 1024×1024, Glow, and StarGAN as its five generative architectures.The architectures are described as well-known state-of-the-art generators.
- Dataset composition: Each architecture has 3,600 training, 2,400 validation, and 2,400 test images per class.The reported splits are defined per class for images generated by each architecture.
- Dataset composition: The experiments use non-overlapping pristine real samples across different architectures.This prevents the selected real samples for different architectures from overlapping.
- Visual examples: The dataset includes representative real and generated images, distinguished in Figure 3 by green and red borders, respectively.Green borders mark real images and red borders mark generated images.
B. Experimental setup
The experimental setup trains XceptionNet on cropped image patches and selects loss parameters using a fixed memory budget and an initial subset of GAN architectures.
- Network training: Networks are trained on randomly cropped 256 × 256 patches, while testing uses central cropped patches.XceptionNet is initialized with ImageNet weights and optimized with ADAM at learning rate 0.001 and batch size 64.
- Parameter selection: The parameter study fixes the memory budget at M = 512 and uses the first three GAN architectures.It evaluates the regularization term and distillation temperature for the proposed methods.
- Parameter selection: The selected settings are T = 2, λ = 1 for MT-MC, λ = 0.5 for MT-SC, and γ = 0.5 in all experiments.These values are reported after the parameter-selection procedure.
- Evaluation condition: Table II reports GAN-image detection accuracy for MT-SC and MT-MC iCaRL after the third GAN is added with memory budget 512.The caption specifies the comparison condition but does not provide the accuracy values here.
C. GAN-detection results
The proposed incremental-learning methods maintain strong GAN-detection accuracy as new architectures are introduced, especially under limited memory. They also extend detection to Style-GAN without affecting prior performance.
- Experimental setup: The experiments feed the network one GAN at a time while retaining only a limited number M of previous samples.The evaluation includes comparisons with state-of-the-art detectors, basic iCaRL, and other incremental strategies.
- Memory-budget evaluation: As memory decreases, the proposed method gains substantially over standard detectors and consistently over basic incremental-learning algorithms.The comparison includes the extreme cases M = ∞ and M = 0, alongside finite memory budgets.
- Increasing GANs: With M = 256, the proposed methods maintain detection accuracy above 90% as the number of considered GAN architectures increases.Accuracy is evaluated jointly on the old classes and the current class; most other techniques degrade as architectures accumulate.
- New architecture: Adding Style-GAN confirms that the proposed methods can increase detector capability without affecting performance on previously learned architectures.The Style-GAN experiment reports accuracy after the last training stage under different memory budgets.
D. GAN-classification results
The proposed system also classifies which GAN architecture generated an image. Classification accuracy exceeds 90% for almost all architectures, with lower accuracy for ProGAN and Glow.
- Architecture classification: The method supports classification of the specific GAN architecture in addition to detecting whether an image is generated.The classification results are shown using the MT-SC method with memory budget M = 256.
- Classification performance: Accuracy is above 90% for almost all architectures but remains above 80% for ProGAN and Glow.The results are presented as a confusion matrix after training on the last GAN.
V. CONCLUSION
The paper addresses GAN-image detection with an incremental-learning strategy designed to handle new architectures without reducing performance on previous ones. It identifies detection of unseen GANs without architecture information as future work.
- Main conclusion: The proposed incremental-learning strategy detects images generated by new GANs without reducing performance on previous ones.The approach builds on a method originally proposed for object classification.
- Future work: Future work will study detection of images generated by new GANs when the generating architecture is unknown.The conclusion identifies this as a more challenging problem than the setting addressed in the paper.