Research map / Adversarial robustness and security / Certified and empirical robustness
SAFARI: Versatile and Efficient Evaluations for Robustness of Interpretability
Wei Huang, Xingyu Zhao, Gaojie Jin, Xiaowei Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use Genetic Algorithm instead of simple Monte Carlo to find worst-case interpretation discrepancy | worst-case evaluation · CIFAR10 (100 seeds) | worst-case MSE: 0.589 -> 42.436 | Improved |
| use Subset Simulation instead of simple Monte Carlo for rare misinterpretation probability | probabilistic evaluation · MNIST F-hat | samples required: 1.13x10^7 -> 15000 | Improved |
| use Subset Simulation on a common event | probabilistic evaluation · CIFAR10 F-hat (ln P=-0.79) | samples required: 2500 vs 2500 | No clear effect |
| increase MH steps in Subset Simulation (ablation) | Subset Simulation · MNIST seed | estimation accuracy | Improved |
| increase GA population size beyond 500 (ablation) | GA · CelebA seed | optimisation objective | No clear effect |
| add input Hessian norm regularisation to training (ablation) | DL model · MNIST | worst-case PCC under F-hat: 0.06 -> 0.17 | Improved |
| add input gradient norm regularisation to training (ablation) | DL model · CIFAR10 | worst-case classification J: 42.58 -> 42.34 | No clear effect |
| add PGD adversarial training (ablation) | DL model · CelebA | worst-case PCC under F-hat: 0.08 -> 0.55 | Improved |
| add PGD adversarial training (ablation) | DL model · CIFAR10 | test accuracy: 0.878 -> 0.703 | Hurt |
| compare XAI methods by robustness to F-hat | DeepLift and Integrated Gradients · MNIST | interpretation robustness | Improved |
Closest papers by meaning
- Towards Trustworthy Deep Learning (AAAI 2024 · similarity 0.89)
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.87)
- Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness (ICLR 2025 · similarity 0.87)
- F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI (ICLR 2025 · similarity 0.86)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.86)
- Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 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.
