Research map / Image generation / Autoregressive and latent generation
STD-FD: Spatio-Temporal Distribution Fitting Deviation for AIGC Forgery Identification
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 STD-FD, a forgery detector for diffusion-generated images that models per-superpixel temporal SSIM sequences of diffusion sampling noise and classifies them with XGBoost on DFactor-based deviation features.
Problem
Detecting AIGC forgeries from diffusion models, where existing reconstruction-error methods treat the generative process as a black box and are sensitive to the pre-trained reconstruction model's domain.
Approach
Decompose the diffusion sampling process into per-superpixel noise-map time series (SSIM between adjacent timesteps), learn discriminative temporal subsequences (DFactors) via a GDTW-based distance and KL-divergence objective, and classify using distribution-fitting deviation features.
Key results
- Achieves average AUC of 94.90% on DeepFaceGen across 12 forgery subsets, with AUC above 90% on all subsets.
- On GenImage cross-generator generalization (Table 1), STD-FD reaches 99.9/91.9 on Midjourney vs DRCT 99.9/89.3 (Score1/Score2 ACC).
- With pre-trained reconstruction models on cat, horse, and bedroom, STD-FD keeps AUC above 92% while DIRE drops to 55-73% AUC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STD-FD instead of DRCT | GenImage cross-generator evaluation · GenImage Midjourney | ACC (Score2): 89.3 -> 91.9 | Helped |
| Use STD-FD instead of DRCT | GenImage cross-generator evaluation · GenImage SDv1.4 | ACC (Score2): 86.6 -> 90.8 | Helped |
| Use STD-FD instead of DRCT | GenImage cross-generator evaluation · GenImage BigGAN | ACC (Score2): 84.0 -> 86.9 | Helped |
| Use STD-FD rather than reconstruction-error methods under a mismatched pretrained reconstruction model | DeepFaceGen forgery detection · DeepFaceGen with cat-pretrained model | AUC: DIRE 73.01; DRCT 74.48; STD-FD 92.78 | Helped |
| Replace Xgb with Logistic Regression (ablation) | STD-FD · GenImage average across subsets | ACC: 90.71 | No clear effect |
| Replace Xgb with Random Forest (ablation) | STD-FD · GenImage average across subsets | ACC: 91.07 | No clear effect |
| Replace Xgb with SVM (ablation) | STD-FD · GenImage average across subsets | ACC: 91.33 | No clear effect |
| Omit superpixel segmentation (ablation) | STD-FD · DeepFaceGen | AUC: 94.90% -> 91.12% | Hurt |
| Apply adversarial PGD attack (ablation) | STD-FD · DeepFaceGen challenging subsets | AUC: 92.53 -> 85.22 | Hurt |
| Vary sampling timestep count among 5, 10, 20, and 50 (ablation) | STD-FD · DeepFaceGen challenging subsets | AUC | No clear effect |
| Apply resizing or JPEG compression (ablation) | STD-FD · DeepFaceGen challenging subsets | AUC | No clear effect |
| Inject adversarial noise during reverse diffusion sampling (ablation) | STD-FD · DeepFaceGen challenging subsets | AUC: -2.5%; -1.8%; -2.0% | Hurt |
Surprising findings
- Adding adversarial noise at each reverse diffusion step changes performance only within about 2.5%.
- Using a larger number of superpixel blocks K does not always improve results; K=10 was the baseline and K=1 dropped only about 1%.
Limitations
- Relies on reconstruction via diffusion sampling, so it requires running the diffusion model per image.
- Performance drops under adversarial attacks (AUC 85-88% on DeepFaceGen subsets versus 92.5% raw).
- Evaluated mostly on GenImage and DeepFaceGen; the authors note end-to-end training is future work.
Experimental setup
Datasets: GenImage DeepFaceGen
Benchmarks: GenImage DeepFaceGen
Metrics: ACC AUC
Baselines: DIRE DRCT Xception EfficientNet-B0 F3-Net RECCE DNADet UnivFD NPR FreqNet CNNSpot GramNet
Builds on: DDIM DIRE DRCT SLIC superpixel GDTW
Closest papers by meaning
- PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals (ICML 2025 · similarity 0.86)
- FakeDiffer: Distributional Disparity Learning on Differentiated Reconstruction for Face Forgery Detection (AAAI 2025 · similarity 0.86)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.85)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.85)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.85)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.85)
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.