Research map / Adversarial robustness and security / Certified and empirical robustness
Interpretability-Guided Test-Time Adversarial Defense
Akshay Kulkarni, Tsui-Wei Weng
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 |
|---|---|---|---|
| add IG-Defense (neuron importance masking with LO-IR) to a base adversarially trained model | TRADES-AWP WRN-34-10 · CIFAR10 | robust accuracy IW-WC (%): 56.12 -> 57.88 | Improved |
| add IG-Defense LO-IR-50 | OAAT ResNet18 · CIFAR100 | robust accuracy IW-WC (%): 27.12 -> 30.12 | Improved |
| add IG-Defense LO-IR-750 on ImageNet-1k | RTB ResNet50 · ImageNet-1k | robust accuracy IW-WC (%): 34.62 -> 35.48 | Improved |
| add IG-Defense CD-IR to LAS model on CIFAR100 | LAS WRN-34-10 · CIFAR100 | clean accuracy (%): 64.89 -> 62.13 | Hurt |
| use existing test-time defenses HD, SODEF, CAAA instead of base model | OAAT ResNet18 · CIFAR10 | robust accuracy IW-WC (%): 51.01 -> 50.68 (HD) | No clear effect |
| use IG-Defense instead of HD (46x cost) | test-time defense · CIFAR10 OAAT | inference time: 46x -> 2x | Improved |
| add randomized smoothing alone (no masking) (ablation) | TRADES-AWP WRN-34-10 · CIFAR10 | robust accuracy IW-WC (%): 56.12 -> 56.14 | No clear effect |
| use two forward passes without masking (ablation) | TRADES-AWP WRN-34-10 · CIFAR10 | robust accuracy IW-WC (%): 56.12 -> 56.12 | No clear effect |
| mask random neurons instead of importance-ranked neurons (ablation) | TRADES-AWP WRN-34-10 · CIFAR10 | robust accuracy IW-WC (%): 56.12 -> 56.15 | No clear effect |
| add neuron importance ranking (CD-IR) masking instead of random masking (ablation) | TRADES-AWP WRN-34-10 · CIFAR10 | robust accuracy IW-WC (%): 56.15 -> 57.46 | Improved |
| add randomized smoothing to the first forward pass of CD-IR (ablation) | IG-Defense · CIFAR10 | robust accuracy IW-WC (%): 56.12 -> 57.46 | Improved |
| mask shallower layers instead of penultimate layer (ablation) | IG-Defense · CIFAR10 ResNet18 | clean and robust accuracy | Hurt |
Closest papers by meaning
- Evaluating the Adversarial Robustness of Adaptive Test-time Defenses (ICML 2022 · similarity 0.88)
- Increasing Confidence in Adversarial Robustness Evaluations (NeurIPS 2022 · similarity 0.87)
- Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM (CVPR 2024 · similarity 0.85)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.85)
- Rethinking the Adversarial Robustness of Multi-Exit Neural Networks in an Attack-Defense Game (CVPR 2025 · similarity 0.85)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 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.
