Research map / Adversarial robustness and security / Certified and empirical robustness
Proper Network Interpretability Helps Adversarial Robustness in Classification
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 |
|---|---|---|---|
| use Int training with 2-class ℓ1 interpretation discrepancy instead of adversarial training | Small · MNIST, PGD ε=0.3 | adversarial test accuracy: 0.890 -> 0.790 | Hurt |
| use Int training with 2-class ℓ1 interpretation discrepancy instead of adversarial training | WResnet · CIFAR-10, PGD ε=8/255 | adversarial test accuracy: 0.170 -> 0.270 | Helped |
| use Int2 training with 2-class ℓ1 interpretation discrepancy | WResnet · CIFAR-10, PGD ε=8/255 | adversarial test accuracy: 0.170 -> 0.290 | Helped |
| use 1-class rather than 2-class interpretation discrepancy (ablation) | Int training · MNIST, PGD ε=0.3 | adversarial test accuracy: 0.790 -> 0.125 | Hurt |
| use 1-class rather than 2-class interpretation discrepancy (ablation) | Int training · CIFAR-10, PGD ε=8/255 | adversarial test accuracy: 0.270 -> 0.065 | Hurt |
| add adversarial loss (ablation) | Int training · MNIST, PGD ε=0.3 | adversarial test accuracy: 0.790 -> 0.855 | Helped |
| add adversarial loss (ablation) | Int training · CIFAR-10, PGD ε=8/255 | adversarial test accuracy: 0.270 -> 0.320 | Helped |
| add adversarial loss (ablation) | Int2 training · MNIST, PGD ε=0.3 | adversarial test accuracy: 0.845 -> 0.840 | No clear effect |
| add adversarial loss (ablation) | Int2 training · CIFAR-10, PGD ε=8/255 | adversarial test accuracy: 0.290 -> 0.335 | Helped |
| promote robust interpretation using Int training | Small · MNIST, AAI ε=0.3 | Kendall's Tau rank correlation: 0.857 -> 0.913 | Helped |
| promote robust interpretation using Int training | WResnet · R-ImageNet, AAI ε=8/255 | Kendall's Tau rank correlation: 0.884 -> 0.946 | Helped |
| use a 2-class rather than 1-class discrepancy measure for ISA (ablation) | interpretability sneaking attack evaluation · MNIST, CAM, ℓ1 | normalized discrepancy score: 3.0723 -> 3.2672 | Helped |
Closest papers by meaning
- Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients (AAAI 2018 · similarity 0.91)
- Interpretation of Neural Networks Is Fragile (AAAI 2019 · similarity 0.90)
- Robust Feature-Level Adversaries are Interpretability Tools (NeurIPS 2022 · similarity 0.89)
- Improved Network Robustness with Adversary Critic (NeurIPS 2018 · similarity 0.89)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.87)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 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.