Research map / Adversarial robustness and security / Certified and empirical robustness
Reliable Evaluation of Attribution Maps in CNNs: A Perturbation-Based Approach
Lars Nieradzik, Henrike Stephani, Janis Keuper
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 |
|---|---|---|---|
| replace pixel masking with FGSM adversarial perturbation in insertion/deletion | attribution map evaluation · Uniform last-place sanity check (15 cases) | cases passed: 11 -> 15 | Improved |
| replace pixel masking with FGSM adversarial perturbation | attribution map evaluation · Canny second-to-last (15 cases) | cases passed: 3 -> 15 | Improved |
| replace deletion with perturbation metric | attribution map evaluation · ranking consistency across 15 combos | Kendall's tau: 0.236 -> 0.466 | Improved |
| replace insertion with perturbation metric | attribution map evaluation · ranking consistency across 15 combos | Kendall's tau: 0.353 -> 0.466 | Improved |
| replace blurred insertion with perturbation metric | attribution map evaluation · ranking consistency | Kendall's tau: 0.432 -> 0.466 | Improved |
| replace insertion/deletion with perturbation metric | attribution map evaluation · monotonicity | monotonicity: 0.671 -> 0.967 | Improved |
| replace insertion/deletion with perturbation metric | attribution map evaluation · smoothness (lower better) | smoothness: 1.658 -> 0.891 | Improved |
| replace blurred insertion with perturbation metric | attribution map evaluation · smoothness (lower better) | smoothness: 1.238 -> 0.891 | Improved |
| use stronger PGD attack instead of FGSM (ablation) | perturbation metric · ChestX-ray8 | attack success rate: 95.80% -> 100% | Improved |
| use stronger PGD attack instead of FGSM (ablation) | perturbation metric · ranking of attribution maps | ranking | No clear effect |
| increase epsilon beyond 1/255 (ablation) | perturbation metric · monotonicity | monotonicity | Hurt |
Closest papers by meaning
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.92)
- Better Understanding Differences in Attribution Methods via Systematic Evaluations (TPAMI 2024 · similarity 0.89)
- Towards Better Understanding Attribution Methods (CVPR 2022 · similarity 0.89)
- Pixel-level Certified Explanations via Randomized Smoothing (ICML 2025 · similarity 0.88)
- One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.86)
- Rethinking Robustness of Model Attributions (AAAI 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.
