Research map / Architectures and efficiency / CNN design
Improved techniques for deterministic l2 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 |
|---|---|---|---|
| Replace the final linear layer with a 1-hidden-layer MLP and train/certify it using CRC-Lip, together with projection pooling (ablation) | LipConvnet · CIFAR-10, LipConvnet-15 | standard accuracy: 77.78% -> 79.44% | Helped |
| Replace the final linear layer with a 1-hidden-layer MLP and train/certify it using CRC-Lip, together with projection pooling (ablation) | LipConvnet · CIFAR-10, LipConvnet-15, radius 36/255 | provable robust accuracy: 62.75% -> 66.99% | Helped |
| Replace the final linear layer with a 1-hidden-layer MLP and train/certify it using CRC-Lip, together with projection pooling (ablation) | LipConvnet · CIFAR-100, LipConvnet-15 | standard accuracy: 48.06% -> 50.79% | Helped |
| Replace the final linear layer with a 1-hidden-layer MLP and train/certify it using CRC-Lip, together with projection pooling (ablation) | LipConvnet · CIFAR-100, LipConvnet-15, radius 36/255 | provable robust accuracy: 34.52% -> 37.50% | Helped |
| Use faster SOC gradient computation (ablation) | LipConvnet-15 · CIFAR-10 | standard accuracy: 77.78% -> 77.75% | No clear effect |
| Use faster SOC gradient computation (ablation) | LipConvnet-15 · CIFAR-10, radius 36/255 | provable robust accuracy: 62.75% -> 62.52% | No clear effect |
| Use faster SOC gradient computation (ablation) | SOC · CIFAR-10, 41 SOC layers | time per epoch: 156.25 -> 106.01; -32.16% | Helped |
| Use projection pooling in place of max pooling (ablation) | LipConvnet · CIFAR-10 and CIFAR-100 | accuracy | Helped |
Closest papers by meaning
- Improved deterministic l2 robustness on CIFAR-10 and CIFAR-100 (ICLR 2022 · similarity 0.96)
- Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks (NeurIPS 2019 · similarity 0.93)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.92)
- Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss (ICML 2025 · similarity 0.92)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.91)
- 1-Lipschitz Layers Compared: Memory Speed and Certifiable Robustness (CVPR 2024 · similarity 0.90)
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.