Research map / Adversarial robustness and security / Attacks on text and graphs
Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition
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 Sharing Ratio Decomposition (SRD), a model-faithful explanation method that decomposes pointwise feature vectors across layers using pre-activation values and sharing ratios, yielding fine-grained, robust saliency maps for CNNs.
Problem
Existing attribution/explanation methods often do not faithfully reflect the model's inference process, are neuron-level and discard inactive-neuron contributions, and are vulnerable to adversarial manipulation of explanations.
Approach
Analyze hidden layers as pointwise feature vectors (channel vectors at each location) instead of scalar neurons, decompose relevance via inner-product sharing ratios computed from pre-activation values (motivated by the Activation-Pattern-Only Prediction observation), and recursively build high-resolution effective receptive fields without modifying model computations.
Key results
- On ResNet50, SRD-input achieves the best Pointing Game (.953) and Stability (.104), and SRD-high achieves the best Attribution Localization (.579) and Fidelity (.142).
- On VGG16, SRD-high achieves the top Pointing Game (.948) and Fidelity (.206), and SRD-input the best Sparseness (.788) and Stability (.099).
- APOP experiments show VGG16 retains 0.575 top-1 accuracy with all inputs replaced by zeros and only the recorded activation pattern kept (original 0.698).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SRD-high instead of the strongest baseline | VGG16 explanations · ImageNet-S50 | Pointing Game: .945 -> .948 | Helped |
| Use SRD-high instead of the strongest baseline | VGG16 explanations · ImageNet-S50 | Fidelity: .199 -> .206 | Helped |
| Use SRD-input instead of the strongest baseline | VGG16 explanations · ImageNet-S50 | Sparseness: .711 -> .788 | Helped |
| Use SRD-input instead of the strongest baseline | VGG16 explanations · ImageNet-S50 | Stability: .100 -> .099 | Helped |
| Use SRD-high instead of the strongest baseline | ResNet50 explanations · ImageNet-S50 | Pointing Game: .946 -> .952 | Helped |
| Use SRD-high instead of the strongest baseline | ResNet50 explanations · ImageNet-S50 | Attribution Localization: .498 -> .579 | Helped |
| Use SRD-high instead of the strongest baseline | ResNet50 explanations · ImageNet-S50 | Fidelity: .135 -> .142 | Helped |
| Use SRD-input instead of the strongest baseline | ResNet50 explanations · ImageNet-S50 | Stability: .112 -> .104 | Helped |
| Analyze with vectors rather than scalar neurons (ablation) | SRD · ResNet50 and VGG16 ablation study | Explanation quality | Helped |
| Use pre-activation rather than post-activation values (ablation) | Vector-based SRD · ResNet50 and VGG16 ablation study | Explanation quality | Helped |
Surprising findings
- Classification accuracy stays well above chance (e.g., 0.575 top-1 for VGG16) when the input is an empty image and only the on/off activation pattern is kept (APOP).
- Post-activation relevance yields overly sparse maps and scalar (neuron-level) relevance yields noisy maps in the ablation study.
Limitations
- Restricted to CNNs; extension to other architectures is claimed but not shown.
- Global and 'how' explanations are only sketched and deferred to future work.
- MLP classifier layers fall back to Grad-CAM, so the method is not purely vector-based end to end.
Experimental setup
Datasets: ImageNet ImageNet-S50 CIFAR-100
Benchmarks: ImageNet-S50
Metrics: Pointing Game Attribution Localization Sparseness Fidelity Stability SSIM PCC
Baselines: Saliency Guided Backprop GradInput InteGrad LRP z+ Smoothgrad Fullgrad GradCAM GradCAM++ ScoreCAM AblationCAM XGradCAM
Builds on: LRP Grad-CAM LayerCAM Effective Receptive Field
Closest papers by meaning
- F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI (ICLR 2025 · similarity 0.86)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.85)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.85)
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.84)
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.84)
- Factor Graph-based Interpretable Neural Networks (ICLR 2025 · similarity 0.84)
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.