Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Binary Models by Pruning Randomly-initialized Networks
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 |
|---|---|---|---|
| Use adaptive pruning (p=0.1, r=0.99) instead of fixed pruning rate (p=1.0, r=0.8) (ablation) | Randomly initialized binary RN34-LBN · CIFAR10 test | robust accuracy: 43.23 -> 44.88 | Helped |
| Add the last batch normalization layer (ablation) | Randomly initialized binary RN34, score initialization a=0.001 · CIFAR10 test | robust accuracy: 33.08 -> 45.06 | Helped |
| Add the last batch normalization layer (ablation) | Randomly initialized binary models · CIFAR10 test | robust accuracy: 40.94 -> 44.65 | Helped |
| Replace Signed Kaiming Constant initialization with binary initialization (ablation) | RN34 with LBN and p=0.1 · CIFAR10 test | robust accuracy: 45.01 -> 45.06 | No clear effect |
| Use binary initialization without the last normalization layer (ablation) | RN34 with adaptive pruning · CIFAR10 test | robust accuracy: 40.94 -> 30.02 | Mixed |
| Use FGSM with ATTA instead of multi-step PGD during training (ablation) | Ours, RN34-LBN · CIFAR10 test | robust accuracy: 45.06 -> 40.77 | Hurt |
| Use FGSM with ATTA instead of multi-step PGD during training (ablation) | Ours, RN34-LBN · CIFAR100 test | robust accuracy: 34.83 -> 34.45 | No clear effect |
| Use adaptive pruning with p=0.1 instead of fixed pruning rate with p=1.0 (ablation) | Randomly initialized binary RN34 · CIFAR10 test | robust accuracy: 42.85 -> 4.35 (r=0.5); 20.56 -> 33.09 (r=0.998) | Mixed |
Closest papers by meaning
- Theoretical evidence for adversarial robustness through randomization (NeurIPS 2019 · similarity 0.89)
- HYDRA: Pruning Adversarially Robust Neural Networks (NeurIPS 2020 · similarity 0.88)
- Drawing Robust Scratch Tickets: Subnetworks with Inborn Robustness Are Found within Randomly Initialized Networks (NeurIPS 2021 · similarity 0.88)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · similarity 0.86)
- Accelerating Adversarially Robust Model Selection for Deep Neural Networks via Racing (AAAI 2024 · similarity 0.86)
- Adversarial Robustness via Deformable Convolution with Stochasticity (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.