Research map / Adversarial robustness and security / Certified and empirical robustness
Does a Hybrid Space-Aware Randomized Defense Improve Empirical and Certified Adversarial Robustness?
Joy Dhar, Manish Kumar Pandey, Behzad Bozorgtabar, Nayyar Zaidi, Wenyu Zhang, Wei-Hong Li, Tingting Mu, Dwarikanath Mahapatra, Mahsa Baktashmotlagh, Trung Le, Chen Chen, Sajib Mistry and 9 more
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 RWAN and SANI hybrid stochasticity over ARS | HySCAN · ImageNet and CIFAR-10 certified, r>0 | certified accuracy: up to 9.6 percentage points | Improved |
| add hybrid space-aware stochasticity over HyCAS | HySCAN · ImageNet and CIFAR-10 certified, r>0 | certified accuracy: up to 1.9 percentage points | Improved |
| add attention-noise hybrid defense over ARS and HyCAS | HySCAN · NIH-CXR / HAM10000 certified (sigma=1.0, r=1.0) | certified accuracy: 34.1 -> 42.9 and 41.4 -> 42.9 (NIH-CXR); 34.6 -> 39.5 and 38.5 -> 39.5 (HAM10000) | Improved |
| raise smoothing noise sigma from 0.25 to 0.50 (ablation) | HySCAN · ImageNet certified, r=2.0 | certified accuracy: 7.15 -> 26.2 | Improved |
| raise smoothing noise sigma from 0.25 to 0.50 (ablation) | HySCAN · ImageNet certified, r=1.0 | certified accuracy: 42.3 -> 42.3 | No clear effect |
| raise smoothing noise sigma from 0.25 to 0.50 (ablation) | HySCAN · CIFAR-10 certified, r=0.75 and r=2.0 | certified accuracy: 45.5 -> 45.8 (r=0.75); 9.94 -> 13.9 (r=2.0) | Mixed |
| use HySCAN instead of CTRW and other empirical defenses | HySCAN · four medical datasets, APGD-20 and AA-20 | robust accuracy: +2.1 to +5.0 percentage points | Improved |
| use HySCAN instead of DCS | HySCAN · EyePACS AA-20 (eps=8/255) | robust accuracy: up to +7.8 points | Improved |
| use HySCAN under strong PGD at eps=0.08 | HySCAN · CIFAR-10 PGD | robust accuracy: about 9 point advantage | Improved |
| use HySCAN with more PGD steps | HySCAN · CIFAR-100 PGD at 100 steps | robust accuracy: about 2 point margin | Improved |
| use HySCAN against HyCAS on clean accuracy | HySCAN · ImageNet certified (sigma=0.25, r=0) | clean accuracy: 72.3 -> 71.9 | Hurt |
Closest papers by meaning
- Adversarial Robustness via Deformable Convolution with Stochasticity (ICML 2025 · similarity 0.88)
- Robust SAM: On the Adversarial Robustness of Vision Foundation Models (AAAI 2025 · similarity 0.84)
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.84)
- ROME is Forged in Adversity: Robust Distilled Datasets via Information Bottleneck (ICML 2025 · similarity 0.84)
- Exploring Adversarial Robustness of Deep State Space Models (NeurIPS 2024 · similarity 0.84)
- Explaining the role of Intrinsic Dimensionality in Adversarial Training (ICML 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.
