Source-linked AI summary
Learning to Detect Fake Face Images in the Wild
Chih-Chung Hsu, Chia-Yen Lee, Yi-Xiu Zhuang
TL;DR
GAN-generated images create security concerns, and binary classifiers may not learn features that generalize across different generators. The paper proposes DeepFD, which uses contrastive loss to learn shared fake-image features before classification. Experiments report successful detection across several state-of-the-art GANs, with stronger performance than baseline methods and localization of unrealistic details.
Problem
Traditional forgery detectors struggle with GAN-generated images, while classifiers trained on limited generators may fail on images from new GANs.
Source: §I · p. 1Method
DeepFD learns jointly discriminative features across GAN-generated and real images with contrastive loss, then concatenates a classifier to detect fake images.
Source: §I · pp. 1–2, A. Jointly Discriminative Feature… · p. 2Results
DeepFD significantly outperforms other techniques in accuracy, precision, and recall, and can localize unrealistic details in generated images.
Source: B. Performance Comparison · p. 4, C. Visualization of the Unrealistic… · p. 4, §IV · p. 4Takeaways & Limitations
Shared cross-GAN features and classifier refinement support generalized fake-image detection while also enabling artifact localization.
Source: B. Performance Comparison · p. 4, §IV · p. 4
Abstract
from arXiv · showhide
Although Generative Adversarial Network (GAN) can be used to generate the realistic image, improper use of these technologies brings hidden concerns. For example, GAN can be used to generate a tampered video for specific people and inappropriate events, creating images that are detrimental to a particular person, and may even affect that personal safety. In this paper, we will develop a deep forgery discriminator (DeepFD) to efficiently and effectively detect the computer-generated images. Directly learning a binary classifier is relatively tricky since it is hard to find the common discriminative features for judging the fake images generated from different GANs. To address this shortcoming, we adopt contrastive loss in seeking the typical features of the synthesized images generated by different GANs and follow by concatenating a classifier to detect such computer-generated images. Experimental results demonstrate that the proposed DeepFD successfully detected 94.7% fake images generated by several state-of-the-art GANs.
I. INTRODUCTION
GAN-generated images create security and social risks, while traditional forgery detectors struggle because generated content lacks the expected intrinsic statistical anomalies. DeepFD addresses this gap by learning shared features across GANs with contrastive loss and adding localization capability.
I. INTRODUCTION
- GAN-generated realistic images can enable fake accounts and other malicious uses affecting social, political, commercial, and personal safety contexts.The motivation includes synthesized high-resolution face images and harmful tampered media.Source: §I · p. 1
- Traditional detectors fail on GAN-generated images because deep networks directly create their content without the unusual intrinsic statistics targeted by earlier methods.Earlier approaches use external signals such as watermarks or intrinsic cues such as sensor noise and compression artifacts.Source: §I · p. 1
- DeepFD learns jointly discriminative features from fake images generated by different GANs using contrastive loss, then classifies computer-generated images.The approach is intended to avoid relying on examples from every possible generator.Source: §I · pp. 1–2
- DeepFD’s fully convolutional architecture can localize unrealistic details in fake images and use those regions to improve detection performance.This localization is presented as a second contribution alongside cross-GAN discrimination.Source: §I · pp. 1–2
II. THE PROPOSED DEEP FORGERY DISCRIMINATOR
DeepFD uses fake images synthesized by multiple GANs together with real images to learn shared features and then classify test images through two connected stages.
II. THE PROPOSED DEEP FORGERY DISCRIMINATOR
- DeepFD first learns jointly discriminative features D1 from fake images synthesized by several GANs and from real images.The training data are collected from one or more example-generative models.Source: §II · p. 2
- The classifier D2 is concatenated with D1 so the resulting network directly distinguishes fake from real images during testing.D2 is connected to the fourth layer of D1.Source: §II · p. 2
A. Jointly Discriminative Feature Learning
The feature-learning stage uses paired images and contrastive loss to learn nonconstant representations that bring same-class examples together while separating different classes.
A. Jointly Discriminative Feature Learning
- Contrastive loss replaces direct feature-distance minimization, which could collapse D1 into a constant mapping.The loss uses paired inputs and a predefined margin.Source: A. Jointly Discriminative Feature… · p. 2
- The proposed pipeline continuously synthesizes training samples with one or more example-generative models for DeepFD training.The flowchart describes the generative models as sources of training examples.Source: Figure 1 · p. 2
- The feature-learning network structure is documented in Table I, whose supplied caption identifies the jointly discriminative feature network.The available table passage provides its scope but not layer-specific values.Source: Table I · p. 2
- For fake-fake or real-real pairs, D1 representations become similar, while the margin term separates pairs from different classes.This pairwise learning captures common characteristics of fake images generated by different GANs.Source: A. Jointly Discriminative Feature… · p. 2
B. Classifier Training
After feature learning, DeepFD adds convolutional and fully connected layers to form an end-to-end classifier trained with cross-entropy and back-propagation.
B. Classifier Training
- DeepFD concatenates a convolutional layer and a fully connected layer to D1, producing an end-to-end architecture for image classification.Alternative classifiers such as SVM and random forest are mentioned but not used in this design.Source: B. Classifier Training · p. 2
- The classifier uses cross-entropy loss and updates D1 parameters during back-propagation so jointly learned features can be refined.This training is intended to improve fake-image detection.Source: B. Classifier Training · pp. 2–3
- Table II compares DeepFD with other methods on accuracy, precision, and recall, and the supplied results passage reports that DeepFD significantly outperforms the baselines.The comparison also reports stronger gains when the training set excludes LSGAN, whose images contain fewer unrealistic details.Source: Table II · p. 3
A. Experimental Settings
The experiments use CelebA faces and synthetic images from five GANs, with comparisons against traditional, classifier, and contrastive-loss ablations.
A. Experimental Settings
- The dataset uses 400,198 training images and 5,000 test images containing real and fake images, with fake images generated by five GANs at 64 × 64 resolution.The five GANs are DCGAN, WGAP, WGAN-GP, LSGAN, and PGGAN; PGGAN images are downsampled to 64 × 64 for fair comparison.Source: A. Experimental Settings · p. 3
- Training uses a 0.5 contrastive-loss margin, Adam optimization, batch size 32, and two initial epochs devoted to pairwise D1 learning.Source: A. Experimental Settings · p. 3
- The study compares traditional forgery detection, DeepFD without contrastive loss, and alternative RFC, SVM, and linear classifiers.Source: A. Experimental Settings · p. 3
B. Performance Comparison
DeepFD uses pairwise contrastive learning to obtain jointly discriminative features, then achieves stronger generalized detection than the compared approaches.
B. Performance Comparison
- Contrastive-loss training makes DeepFD converge easily and reach higher performance than the corresponding model without contrastive loss.The comparison is reported for a training set excluding LSGAN.Source: B. Performance Comparison · p. 4
- DeepFD significantly outperforms other techniques in accuracy, precision, and recall when one GAN is excluded from training.The result supports greater generalization to fake images from an unseen GAN.Source: B. Performance Comparison · p. 4
- The larger gain without LSGAN is attributed to LSGAN producing fewer unrealistic details, which limits purely supervised learning of common fake-image features.Source: B. Performance Comparison · p. 4
C. Visualization of the Unrealistic Details in the Fake Images
DeepFD’s fully convolutional design enables feature-map visualization to localize unrealistic details in generated faces.
C. Visualization of the Unrealistic Details in the Fake Images
- Most artifacts in six PGGAN examples are successfully localized through high-activation feature representations in the final classifier layer.The visualized regions can be investigated to further improve detection performance.Source: C. Visualization of the Unrealistic… · p. 4
IV. CONCLUSIONS
The paper concludes that DeepFD combines contrastive learning and classifier refinement to detect GAN-generated images while also exposing unrealistic regions.
IV. CONCLUSIONS
- DeepFD uses contrastive loss to capture jointly discriminative features across fake images generated by different GANs.Source: §IV · p. 4
- Classifier refinement further boosts classification performance and supports visualization of unrealistic details in fake images.Source: §IV · p. 4, Figure 4 · p. 4