Research map / Adversarial robustness and security / Certified and empirical robustness
Feature Separation and Recalibration for Adversarial Robustness
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 |
|---|---|---|---|
| Add FSR to AT | ResNet-18 · CIFAR-10 | PGD-20 accuracy (%): 48.22 -> 52.47 | Helped |
| Add FSR to TRADES | ResNet-18 · SVHN | PGD-20 accuracy (%): 47.12 -> 51.49 | Helped |
| Replace deactivation methods with FSR | AT, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): AT 45.51; FD 45.82; CAS 46.46; CIFS 47.26; FSR 48.34 | Helped |
| Add Recalibration to the separated non-robust feature (ablation) | FSR, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): 33.11 -> 46.93 | Helped |
| Combine robust and recalibrated non-robust features (ablation) | FSR, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): 47.89 -> 48.34 | Helped |
| Add Separation before Recalibration (ablation) | FSR with AT, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): 47.60 -> 48.34 | Helped |
| Add Recalibration to the robust-only feature path (ablation) | FSR with AT, ResNet-18 · CIFAR-10 | AutoAttack accuracy (%): 44.96 -> 46.41 | Helped |
| Use a differentiable Gumbel mask instead of a binary mask (ablation) | FSR, ResNet-18 · CIFAR-10 | AutoAttack accuracy (%): 44.26 -> 46.41 | Helped |
| Use learned Separation Net scores instead of greedy or random masks (ablation) | FSR, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): Greedy 46.42; Random 46.77; w/o Separation 47.60; w/ Separation 48.34 | Helped |
| Place FSR after Block4 rather than earlier blocks (ablation) | ResNet-18 with AT · CIFAR-10 | Ensemble accuracy (%): Block1 44.89; Block2 45.28; Block3 46.33; Block4 48.34 | Helped |
| Use the highest-scoring mispredicted class as the separation target (ablation) | FSR, ResNet-18 · CIFAR-10 | Ensemble accuracy (%): Uniform 47.90; Entropy max. 47.88; Avg. targeted loss 47.42; Mispredicted 48.34 | Helped |
| Add FSR to adversarial training | ResNet-18 · CIFAR-100 | Ensemble accuracy (%): AT 22.46 -> AT + FSR 22.95 | Helped |
Closest papers by meaning
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
- Robustness Reprogramming for Representation Learning (ICLR 2025 · similarity 0.86)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.86)
- Mitigating Feature Gap for Adversarial Robustness by Feature Disentanglement (AAAI 2025 · similarity 0.86)
- Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM (CVPR 2024 · similarity 0.85)
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.