Research map / Image generation / Autoregressive and latent generation
Diffusion Models for Attribution
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 a diffusion-model-based information bottleneck method that learns per-pixel noise levels (via MMSE/SNR connection) to produce high-resolution pixel-level attribution maps for image classifiers.
Problem
Generating accurate, high-resolution pixel-level attribution maps for explaining image classifier decisions, avoiding gradient vanishing, blurred maps, and the cost of black-box perturbation methods; mutual information in pixel space is intractable.
Approach
Treats diffusion noise as an information bottleneck: links mutual information to MMSE via the I-MMSE relation so that minimizing per-pixel SNR reduces mutual information, and trains a U-Net to predict per-pixel noise scales that preserve classifier confidence while minimizing SNR, then integrates MMSE to get attribution.
Key results
- Reaches 90% confidence with about 2-3% of pixel channel values, versus at least 13% for other methods (MINM ratio 0.025 vs 0.132 for InputIBA and 0.165 for IBA).
- Achieves the best DAUC of 0.854 vs 0.841 for InputIBA and 0.774 for IBA.
- Achieves the best segmentation-based ratio of 0.511 vs 0.488 for IBA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use diffusion-based pixel-level attribution with an information bottleneck | VGG16 explanations · ImageNet (1,000 images) | MINM: 0.132 ± 0.002 -> 0.025 ± 0.005 | Helped |
| Use diffusion-based pixel-level attribution with an information bottleneck | VGG16 explanations · ImageNet (1,000 images) | DAUCs: 0.841 ± 0.002 -> 0.854 ± 0.005 | Helped |
| Use diffusion-based pixel-level attribution with an information bottleneck | VGG16 explanations · ImageNet bounding boxes | Box-Ratio: 0.998 ± 0.001 -> 0.998 ± 0.000 | No clear effect |
| Use diffusion-based pixel-level attribution with an information bottleneck | VGG16 explanations · FSS-1000 semantic segmentation | SR: 0.488 ± 0.004 -> 0.511 ± 0.002 | Helped |
Surprising findings
- Methods such as Integrated Gradients, Guided-BP and DeepLIFT score near 0.99 on MINM, indicating they need almost all pixels to reach the confidence threshold.
Limitations
- Relies on integrating MMSE via trapezoidal rule, which requires numerical approximation.
- Evaluation uses a single main model (VGG16) in the main text; other architectures are only in the appendix.
Experimental setup
Datasets: ImageNet FSS-1000
Benchmarks: ImageNet FSS-1000
Metrics: MNI/MINM pixel ratio Insertion-Deletion DAUC Box-Ratio Segmentation-based ratio (SR) SSIM parameter randomization sanity check
Baselines: IBA InputIBA Integrated Gradients Guided-Backprop DeepLIFT HSIC-Attribution
Builds on: Information Bottleneck for Attribution (IBA) InputIBA DDPM HSIC-Attribution I-MMSE relation
Closest papers by meaning
- Diffusion Attribution Score: Evaluating Training Data Influence in Diffusion Models (ICLR 2025 · similarity 0.88)
- Intriguing Properties of Data Attribution on Diffusion Models (ICLR 2024 · similarity 0.87)
- Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers (CVPR 2025 · similarity 0.86)
- Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew (NeurIPS 2026 · similarity 0.86)
- Influence Functions for Scalable Data Attribution in Diffusion Models (ICLR 2025 · similarity 0.86)
- Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model (NeurIPS 2024 · similarity 0.86)
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.