MLEP: Multi-granularity Local Entropy Patterns for Universal AI-generated Image Detection

TL;DR

The paper tackles the problem of robust AI-generated image detection across diverse generation models and scenes. It introduces Multi-granularity Local Entropy Patterns (MLEP), which builds entropy-based feature maps from fine-grained patch shuffling and multi-scale resampling to suppress semantic content and highlight pixel relationships. A CNN classifier trained on LEP features achieves superior generalization, outperforming state-of-the-art methods across 32 generative models (GANs and diffusion-based) in open-world evaluations. The work demonstrates the practical potential of entropy-based cues for content-agnostic AIGI detection and points to future improvements in adaptive entropy computation and robustness to common image processing operations.

Abstract

Advancements in image generation technologies have raised significant concerns about their potential misuse, such as producing misinformation and deepfakes. Therefore, there is an urgent need for effective methods to detect AI-generated images (AIGI). Despite progress in AIGI detection, achieving reliable performance across diverse generation models and scenes remains challenging due to the lack of source-invariant features and limited generalization capabilities in existing methods. In this work, we explore the potential of using image entropy as a cue for AIGI detection and propose Multi-granularity Local Entropy Patterns (MLEP), a set of entropy feature maps computed across shuffled small patches over multiple image scaled. MLEP comprehensively captures pixel relationships across dimensions and scales while significantly disrupting image semantics, reducing potential content bias. Leveraging MLEP, a robust CNN-based classifier for AIGI detection can be trained. Extensive experiments conducted in an open-world scenario, evaluating images synthesized by 32 distinct generative models, demonstrate significant improvements over state-of-the-art methods in both accuracy and generalization.

Figures (6)

• Figure 1: Comparison of local entropy distributions between real and AI-generated images using $2 \times 2$ patches, with entropy values from $\{0, 0.8, 1.0, 1.5, 2.0\}$. Real images consistently show a higher likelihood of entropy reaching 2.0.
• Figure 2: The MLEP framework, composed of three steps: Patch Shuffling, Multi-Scale Resampling, and Local Entropy Patterns computation.
• Figure 3: Illustration of LEP computation in a single patch (a) and its intra-block and inter-block views (b) when using a $2\times 2$ sliding window with stride of 1 across shuffled image patches.
• Figure 4: Visualization of local entropy patterns for several real–fake image pairs, along with their differences in the pixel, entropy, and Fourier domains.
• Figure 5: Qualitative comparison among Zheng zheng2024breaking, NPR tan2024rethinking, and our method. LEP preserves minimal visible semantics, while MLEP (without resampling) further suppresses semantic content.