Research map / Adversarial robustness and security / Certified and empirical robustness
On the Robustness of Removal-Based Feature Attributions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train neural networks with increasing weight decay | Removal-based feature attributions · UCI wine quality, MNIST, CIFAR-10, Imagenette | attribution difference under input perturbations | Helped |
| Use conditional rather than baseline or marginal feature removal | Shapley attributions · Synthetic data, model perturbations | attribution difference between perturbed models | Helped |
| Use conditional rather than baseline or marginal feature removal | Removal-based attributions · Input perturbations | attribution robustness | Hurt |
| Increase correlation between features | Conditional feature removal · Synthetic data, input perturbations | attribution difference | Hurt |
| Increase the model Lipschitz constant by scaling logistic regression coefficients | Shapley attributions · Synthetic data, input perturbations | slope of attribution difference versus perturbation norm | Hurt |
| Randomize network parameters in cascading order | Removal-based attributions · Wine quality, MNIST, CIFAR-10, Imagenette | similarity to original attributions | Hurt |
| Randomize all network parameters | Integrated Gradients and Grad × Input · Wine quality, MNIST | similarity to original attributions: around $0.5$ | No clear effect |
| Use Banzhaf rather than other bounded, symmetric linear summary methods | Removal-based feature attributions · Theoretical robustness analysis | spectral norm of summary operator: $1/2^{d/2 - 1}$ | Helped |
| Use Shapley rather than other bounded, efficient linear summary methods | Removal-based feature attributions · Theoretical robustness analysis | spectral norm of summary operator: $\sqrt{2/d}$ | Helped |
Closest papers by meaning
- Rethinking Robustness of Model Attributions (AAAI 2024 · similarity 0.91)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.87)
- Data-faithful Feature Attribution: Mitigating Unobservable Confounders via Instrumental Variables (NeurIPS 2024 · similarity 0.87)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.86)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.86)
- Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining (NeurIPS 2025 · 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.