[Paper Review] A Single Simple Patch is All You Need for AI-generated Image Detection

Jiaxuan Chen, Jieteng Yao+1 more|arXiv (Cornell University)|Feb 2, 2024

COVID-19 diagnosis using AIMedicine3 citations

TL;DR

This paper proposes SSP (Single Simple Patch), a highly effective and generalizable method for detecting AI-generated images by extracting a single, minimal 32×32 patch from an image and analyzing its noise pattern via a binary classifier. Despite its simplicity, SSP achieves a 14.6% relative improvement over the strongest baseline on the GenImage dataset, demonstrating superior cross-generator generalization across diverse diffusion and GAN-based generators.

ABSTRACT

The recent development of generative models unleashes the potential of generating hyper-realistic fake images. To prevent the malicious usage of fake images, AI-generated image detection aims to distinguish fake images from real images. However, existing method suffer from severe performance drop when detecting images generated by unseen generators. We find that generative models tend to focus on generating the patches with rich textures to make the images more realistic while neglecting the hidden noise caused by camera capture present in simple patches. In this paper, we propose to exploit the noise pattern of a single simple patch to identify fake images. Furthermore, due to the performance decline when handling low-quality generated images, we introduce an enhancement module and a perception module to remove the interfering information. Extensive experiments demonstrate that our method can achieve state-of-the-art performance on public benchmarks.

Motivation & Objective

• To address the poor generalization of existing AI-generated image detection methods across diverse generative models, especially across different diffusion and GAN-based generators.
• To develop a simple yet robust baseline method that achieves high performance without complex architectures or extensive hyperparameter tuning.
• To demonstrate that minimal image patches can contain sufficient discriminative noise patterns for reliable fake image detection.

Proposed method

• The method extracts 192 random 32×32 patches from each input image and selects the one with the lowest texture complexity as the 'simplest patch' to minimize structural content and maximize noise visibility.
• The noise pattern of the selected patch is extracted using standard high-pass filters (e.g., SRM filters) to emphasize fine-scale artifacts typical in synthetic images.
• The resulting noise representation is fed into a binary classifier, which distinguishes real from fake images based on subtle statistical differences in noise patterns.
• The approach relies on a minimal pipeline: patch selection, noise extraction via simple filters, and classification—avoiding complex feature engineering or multi-branch networks.
• The method is trained on one generator’s training subset and evaluated on test subsets from other generators, enabling cross-generator generalization evaluation.

Experimental results

Research questions

• RQ1 Can a single, minimally processed image patch achieve competitive performance in detecting AI-generated images across diverse generators?
• RQ2 How does the performance of a simple patch-based method compare to complex state-of-the-art models in cross-generator detection scenarios?
• RQ3 Does the noise pattern in a single simple patch contain sufficient discriminative information to distinguish between real and AI-generated images?
• RQ4 To what extent does the proposed method generalize across different generative models, including both diffusion and GAN-based models?

Key findings

• SSP achieves an average accuracy of 93.5% on the GenImage test set, representing a 14.6% relative improvement over the strongest baseline (GenDet) at 81.6%.
• On hard cross-generator settings such as SD V1.4 → VQDM, SSP achieves 94.4% accuracy, significantly outperforming baselines like CNNSpot (56.7%) and F3Net (62.1%).
• In cross-generator evaluations, SSP consistently outperforms ResNet-50 and other strong baselines across all 64 settings, demonstrating robust generalization.
• The method achieves near-perfect performance (99.4%–99.9%) on within-generator detection for Midjourney and Stable Diffusion, indicating strong detection capability on familiar generators.
• Even with minimal architectural complexity, SSP generalizes effectively across diverse generators, including BigGAN, ADM, and GLIDE, where baselines often fail.