Detection of synthetic or AI-generated images is an increasingly important challenge as generative models (GANs, diffusion models, text-to-image systems) improve in quality. Methods span frequency-domain analysis, forensic feature extraction, and deep learning approaches trained to distinguish real from synthetic pixels.
A key challenge is generalization: detectors trained on images from one generator often fail on images from a different generator, requiring either robust feature spaces or continuous model retraining as new synthesis methods emerge.
Key papers
- 2023-ojha-universal-fake-image-detectors — Demonstrates generalization failure in conventional real-vs-fake classifiers trained on one generator family; proposes using pre-trained CLIP-ViT features with simple nearest-neighbor or linear classification; achieves strong performance on unseen diffusion models, DALL-E, and autoregressive generators without generator-specific training.
- 2020-frank-frequency-gan-detection — Frequency-domain analysis showing GAN artifacts (grid patterns from upsampling) visible in DCT spectrum; simple linear classifiers achieve 100% accuracy; robust to spatial perturbations.
- 2019-wang-cnn-generated-images — Cross-generator generalization study: trains on ProGAN, evaluates on StyleGAN, BigGAN, CycleGAN, DeepFakes; achieves 87–99.6% AP; frequency artifacts as generalization cue.
- 2026-basu-swin-efficientnet-deepfake-forensics — Hybrid EfficientNet-B0 + Swin-Transformer for resource-efficient GAN face detection; 99% accuracy on balanced real/fake dataset.
- 2020-tolosana-deepfakes-survey — Comprehensive review of GAN-based face synthesis and detection methods; covers StyleGAN, ProGAN, and evaluation datasets.
Related topics
- deepfake-detection (facial image/video synthesis detection, a subset)
- generative-models (GAN and diffusion model architectures)
- media-forensics (broader forensic analysis field)
- face-forensics (facial image manipulation detection)