Research map / Image generation / Autoregressive and latent generation
MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MLEP, entropy-based feature maps computed over shuffled small patches at multiple resampling scales, fed to a ResNet-50 classifier to detect AI-generated images across 32 generators.
Problem
Detecting AI-generated images that generalize across diverse generative models and content types, given the lack of source-invariant features in existing detectors.
Approach
Uses local Shannon entropy over 2x2 sliding windows (instead of NPR's pixel differences) on finely shuffled patches and multi-scale resampled images, so the resulting entropy maps capture pixel randomness while suppressing semantic content.
Key results
- MLEP achieves 97.1% mean accuracy and 98.9% mean A.P. across all 32 test generators, versus 94.0% accuracy for NPR and 91.9% for FatFormer.
- On the Diffusion-Set, MLEP reaches 97.8% accuracy and 99.6% A.P., while being trained only on ProGAN data.
- Ablation: patch shuffling contributes more than multi-scale resampling; patch size 2 and scales {1,1/2,1/4} work best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add patch shuffling to LEP (ablation) | MLEP · Mean | Acc.: 94.6 -> 96.6 | Helped |
| Add multi-scale resampling to LEP with patch shuffling (ablation) | MLEP · Mean | Acc.: 96.6 -> 97.1 | Helped |
| Add patch shuffling to LEP without multi-scale resampling (ablation) | MLEP · Mean | Acc.: 94.3 -> 96.6 | Helped |
| Use patch size l=2 rather than l=4 or l=8 (ablation) | MLEP · Mean | Acc.: 95.5 -> 97.1; 94.9 -> 97.1 | Helped |
| Use scales {1, 1/2, 1/4} rather than {1, 1/2} (ablation) | MLEP with l=2 · Mean | Acc.: 96.6 -> 97.1 | Helped |
| Add scale 1/8 to {1, 1/2, 1/4} (ablation) | MLEP with l=2 · Mean | Acc.: 97.1 -> 93.5 | Hurt |
| Replace bilinear interpolation with bicubic (ablation) | MLEP · Mean | Acc.: 97.1 -> 96.9 | No clear effect |
| Replace bilinear interpolation with nearest-neighbor (ablation) | MLEP · Mean | Acc.: 97.1 -> 95.7 | Hurt |
| Increase LEP sliding-window stride from 1 to 2 (ablation) | MLEP · Mean | Acc.: 97.1 -> 95.9 | Hurt |
| Replace ResNet-50 with ResNet-18 or ResNet-34 (ablation) | MLEP · Mean | Acc.: 97.1 -> 96.8; 97.1 -> 96.9 | Hurt |
| Replace ResNet-50 with ResNet-101 (ablation) | MLEP · Mean | Acc.: 97.1 -> 97.1 | No clear effect |
| Use MLEP instead of NPR | AIGI detection · Mean | Acc.: 94.0 -> 97.1 | Helped |
Surprising findings
- Bilinear and bicubic resampling outperform nearest-neighbor, attributed to richer entropy variation.
- Adding scale 1/8 degrades GAN-set accuracy from 96.4% to 91.7% at patch size 2.
Limitations
- Entropy is computed only within 2x2 windows because complexity grows exponentially with larger windows.
- Robustness to common processing operations (compression, blurring, noise) is not yet studied.
Experimental setup
Datasets: ForenSynths LSUN GANGen-Detection DiffusionForensics UniversalFakeDetect Discord-collected Midjourney and DALL·E 2 images
Benchmarks: ForenSynths GAN-Set (16 GAN models) Diffusion-Set (16 diffusion models)
Metrics: accuracy average precision
Baselines: CNNDet F3Net LGrad UnivFD CLIPping NPR Zheng FreqNet FatFormer
Builds on: NPR Zheng et al. patch shuffling detector Shannon information entropy
Closest papers by meaning
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.85)
- Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection (ICML 2025 · similarity 0.84)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.83)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.83)
- Are High-Quality AI-Generated Images More Difficult for Models to Detect? (ICML 2025 · similarity 0.82)
- Learning by Erasing: Conditional Entropy Based Transferable Out-of-Distribution Detection (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.