Research map / Adversarial robustness and security / Certified and empirical robustness
Training Adversarially Robust Sparse Networks via Bayesian Connectivity Sampling
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 |
|---|---|---|---|
| Bayesian connectivity sampling (dynamic rewiring) vs static random sparse connectivity from scratch | sparse VGG-16 with robust training · CIFAR-10 99% sparsity | robust accuracy (%): 42.1 vs 27.0 (Standard AT); 42.3 vs 32.2 (RST) | Helped |
| Dynamic rewiring with sparse connectivity from scratch vs fixed learned per-layer connectivity (ablation) | sparse ResNet-18 / WideResNet-28-4 with robust objectives · CIFAR-100 and SVHN 99% sparsity | robust accuracy (%): 22.2 vs 19.3 (CIFAR-100 TRADES); 52.7 vs 48.3 (SVHN PGD50) | Helped |
| Sparse end-to-end robust training instead of robustness-aware pruning of dense pretrained networks (HYDRA) | VGG-16 at 90% sparsity, CIFAR-10 RST · CIFAR-10 AA∞ | AA∞ robust accuracy (%): 44.98 vs 45.46 (-0.48) | No clear effect |
| Sparse end-to-end robust training instead of HYDRA pruning | WideResNet-28-4 at 90% sparsity, CIFAR-10 RST · CIFAR-10 AA∞ | AA∞ robust accuracy (%): 49.78 vs 51.74 (-1.96) | Hurt |
| Sparse end-to-end robust training instead of HYDRA pruning | WideResNet-28-4 at 99% sparsity, SVHN TRADES · SVHN AA∞ | AA∞ robust accuracy (%): 40.78 vs 38.80 (+1.98) | Helped |
| Sparse end-to-end robust training instead of HYDRA pruning at 99% sparsity for standard AT (ResNet-18) | ResNet-18 standard AT CIFAR-10 99% sparsity · CIFAR-10 robust accuracy | robust accuracy (%): 44.8 vs 41.6 (HYDRA) | Helped |
| Sparse end-to-end robust training with random static connectivity baseline at 99% sparsity under natural training | VGG-16 natural training · CIFAR-10 99% sparsity | clean accuracy (%): 87.7 vs 56.9 (Random) | Helped |
Closest papers by meaning
- Learning Adversarially Robust Sparse Networks via Weight Reparameterization (AAAI 2023 · similarity 0.89)
- Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · similarity 0.86)
- Leveraging Low-Rank and Sparse Recurrent Connectivity for Robust Closed-Loop Control (ICLR 2024 · similarity 0.86)
- Enhancing Adversarial Robustness in SNNs with Sparse Gradients (ICML 2024 · similarity 0.86)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
- SAFE: Finding Sparse and Flat Minima to Improve Pruning (ICML 2025 · 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.