Research map / Adversarial robustness and security / Certified and empirical robustness
Don’t Teach Instability, Teach Robustness: Selective Sensitivity Gating for Adversarial Robust Distillation
Jingqi Ji, Quan Kong, Chaojie Gu, Yuanchao Shu, Cong Wang
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 Semantic Logit Gate to AdaAD baseline (ablation) | AdaAD · CIFAR-10 ResNet18 | AutoAttack acc: 51.41 -> 53.11 | Improved |
| Add target-class gradient robust distillation (TRD) (ablation) | AdaAD · CIFAR-10 ResNet18 | PGD-20 acc: 55.43 -> 57.66 | Improved |
| Add non-target-class gradient distillation (NTRD) (ablation) | AdaAD · CIFAR-10 ResNet18 | AutoAttack acc: 51.41 -> 52.96 | Improved |
| Add TRD and NTRD on top of SLG (full AEGIS) (ablation) | AEGIS · CIFAR-10 ResNet18 | AutoAttack acc: 53.35 (SLG+TRD) vs 53.22 (full) | No clear effect |
| AEGIS vs ARD | ResNet18 · CIFAR-10 | AutoAttack acc: 50.10 -> 53.22 | Improved |
| AEGIS vs DGAD on clean accuracy | ResNet18 · CIFAR-10 | Clean acc: 85.46 -> 84.69 | Hurt |
| Replace SLG with discard or ground-truth label variants (ablation) | SLG · CIFAR-10 | AutoAttack acc | Improved |
| Use selective SCG instead of naive gradient alignment (ablation) | AEGIS · CIFAR-10 | PGD-20 acc: 57.39 -> 57.89 | Improved |
| Tune beta=0.7 and alpha=2.0 (ablation) | AEGIS · CIFAR-10 | PGD acc | Mixed |
| Condition logit matching on teacher reliability (ablation) | AEGIS · 2K CIFAR-10 test samples | student-superior samples: 22.19% -> 97.75% | Improved |
Closest papers by meaning
- CIARD: Cyclic Iterative Adversarial Robustness Distillation (ICCV 2025 · similarity 0.92)
- GCMRD: Global Consistency Multi-teacher Robustness Distillation (ECCV 2026 · similarity 0.91)
- Indirect Gradient Matching for Adversarial Robust Distillation (ICLR 2025 · similarity 0.89)
- Out of Thin Air: Exploring Data-Free Adversarial Robustness Distillation (AAAI 2024 · similarity 0.89)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.89)
- Robust Distillation via Untargeted and Targeted Intermediate Adversarial Samples (CVPR 2024 · similarity 0.88)
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.
