LOTA: Bit-Planes Guided AI-Generated Image Detection

TL;DR

LOTA tackles the challenge of distinguishing AI-generated images from real ones by exploiting intrinsic noise patterns in the lowest bit-planes. It combines Bit-Planes Guided Noisy Image Generation, a Maximum Gradient Patch Selection strategy, and two lightweight classifiers (NBC and NGC) to produce a fast, robust detector that generalizes across generators. On GenImage, LOTA achieves high accuracy (≈98.9%) with millisecond inference and a compact model (~23.6M parameters), outperforming prior error-based methods and showing strong cross-generator resilience. This approach offers a practical, efficient alternative for AI-generated image detection with broad applicability in security and content verification contexts.

Abstract

The rapid advancement of GAN and Diffusion models makes it more difficult to distinguish AI-generated images from real ones. Recent studies often use image-based reconstruction errors as an important feature for determining whether an image is AI-generated. However, these approaches typically incur high computational costs and also fail to capture intrinsic noisy features present in the raw images. To solve these problems, we innovatively refine error extraction by using bit-plane-based image processing, as lower bit planes indeed represent noise patterns in images. We introduce an effective bit-planes guided noisy image generation and exploit various image normalization strategies, including scaling and thresholding. Then, to amplify the noise signal for easier AI-generated image detection, we design a maximum gradient patch selection that applies multi-directional gradients to compute the noise score and selects the region with the highest score. Finally, we propose a lightweight and effective classification head and explore two different structures: noise-based classifier and noise-guided classifier. Extensive experiments on the GenImage benchmark demonstrate the outstanding performance of our method, which achieves an average accuracy of \textbf{98.9\%} (\textbf{11.9}\%~$\uparrow$) and shows excellent cross-generator generalization capability. Particularly, our method achieves an accuracy of over 98.2\% from GAN to Diffusion and over 99.2\% from Diffusion to GAN. Moreover, it performs error extraction at the millisecond level, nearly a hundred times faster than existing methods. The code is at https://github.com/hongsong-wang/LOTA.

Figures (6)

• Figure 1: Comparison of least bit-planes between real images and AI-generated images. We extract the 1st, 2nd and 3rd least bit-planes from both types of images, separately. We find that images of low bit-planes of real images and AI-generated images have different noise patterns and distributions that can be used for distinguishing between them. Low bit-planes of fake images contain artifacts that are invisible in RGB images.
• Figure 2: Overview of our method. First, we decompose the image into 8 bit-planes, and compose least bit-planes to generate the noise representation. Second, we crop the noise image into several patches, and select the patch with the highest gradient-based score. Finally, a classification head is applied.
• Figure 3: Visualizations of generated noisy images by scaling and thresholding in the BGNIG module. We compare corresponding noise images for real and AI-generated images. Both scaling and thresholding methods effectively extract the noise patterns of images. The brightness distribution of noisy images from real images is relatively regular. However, noisy images from synthetic images contain several regions with artifacts.
• Figure 4: Structure of the classifier. Two different classifiers are applied. For the noise-based classifier, the low-bit patch is directly fed into the ResNet. For the noise-guided classifier, the low-bit patch and the feature map of raw images are combined by using the noise-guided multi-head attention.
• Figure 5: Comparison of cross-generator generalization with existing methods. DIRE wang2023dire, LaRE$^2$luo2024lare and ESSP chen2024single are selected as comparison methods. These models and ours are trained on eight training subsets and tested on eight testing subsets in the GenImage. The noise-based classifier is applied on both training and testing. The vertical axis represents the training subsets, and the horizontal axis represents the test subsets. Darker colors indicate higher accuracy.