SFLD: Reducing the content bias for AI-generated Image Detection

TL;DR

This paper tackles AI-generated image detection by addressing generalization gaps across unseen generators and degradation-prone fingerprints. It introduces SFLD, a patch-shuffling detector that ensembles classifiers trained on multiple patch sizes to fuse high-level semantic cues with low-level texture artifacts, using a CLIP-ViT-L/14 backbone. It also introduces TwinSynths, a benchmark of visually near-identical real–fake image pairs to evaluate content preservation and generalization. Empirical results show SFLD achieves state-of-the-art AP on conventional benchmarks, maintains high performance on unseen generators, and demonstrates robustness to Gaussian blur and JPEG degradation, with strong results on low-level vision/perceptual benchmarks. The work provides a practical detector with broad generalization and a benchmarking framework for realistic evaluation in real-world settings.

Abstract

Identifying AI-generated content is critical for the safe and ethical use of generative AI. Recent research has focused on developing detectors that generalize to unknown generators, with popular methods relying either on high-level features or low-level fingerprints. However, these methods have clear limitations: biased towards unseen content, or vulnerable to common image degradations, such as JPEG compression. To address these issues, we propose a novel approach, SFLD, which incorporates PatchShuffle to integrate high-level semantic and low-level textural information. SFLD applies PatchShuffle at multiple levels, improving robustness and generalization across various generative models. Additionally, current benchmarks face challenges such as low image quality, insufficient content preservation, and limited class diversity. In response, we introduce TwinSynths, a new benchmark generation methodology that constructs visually near-identical pairs of real and synthetic images to ensure high quality and content preservation. Our extensive experiments and analysis show that SFLD outperforms existing methods on detecting a wide variety of fake images sourced from GANs, diffusion models, and TwinSynths, demonstrating the state-of-the-art performance and generalization capabilities to novel generative models.

Figures (16)

• Figure 1: Class-wise detection accuracy for StyleGAN-{bedroom, car, cat} class categories. The bedroom class does not appear at training, while car and cat does. UnivFDojha2023towards catastrophically fails to identify synthetic bedroom images, which hints at model bias towards high-level image content.
• Figure 2: Comparison of benchmarks. (a) Real images and fake GAN images are sampled from the test ProGAN set in the ForenSynthswang2020cnn. Fake diffusion images are sampled from benchmark of Ojha et al.ojha2023towards, each from LDM, GLIDE and DALL-E dataset. (b) Real images are sampled from ImageNet dataset, and corresponding fake images are generated by each model.
• Figure 3: Architecture of the proposed fake image detector (SFLD). $z_{s_i}$ refers to the logit score generated from an input image processed via $s_i\text{×}s_i$ patch size. $\Sigma$ indicates weighted sum.
• Figure 4: Scatter plots of per-sample scores. X-axis is the UnivFD logits, and Y-axis is the logit from PatchShuffle with patch size 28. The decision boundary of UnivFD (red) and SFLD (green) are shown. See \ref{['sec:additional-results-on-scatter-plots']} for extended results.
• Figure 5: A closer look into the failure cases from the StyleGAN-generated test images.