Research map / Adversarial robustness and security / Certified and empirical robustness
Robust-by-Design Classification via Unitary-Gradient Neural Networks
Fabio Brau, Giulio Rossolini, Alessandro Biondi, Giorgio Buttazzo
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 |
|---|---|---|---|
| Use unbounded rather than bounded UPD (ablation) | UGNN · CIFAR10 test | accuracy: UGNN+Abs: 71.9±0.29 -> 72.1±0.54 (Std.Norm); 69.2±0.31 -> 68.9±0.81 (Raw). UGNN+MaxMin: 72.6±0.79 -> 72.7±0.38 (Std.Norm | Mixed |
| Apply standard normalization to inputs (ablation) | UGNN · CIFAR10 test | accuracy: UGNN+Abs+updB: 69.2±0.31 -> 71.9±0.29; UGNN+MaxMin+updB: 70.4±0.52 -> 72.6±0.79; UGNN+OPLU+updB: 70.5±0.39 -> 71.9±0.09 | Mixed |
| Replace MaxMin or OPLU with abs activation (ablation) | UGNN · CIFAR10 test | accuracy: Std.Norm: 71.9±0.09 to 72.1±0.54 (OPLU to Abs, updU); Raw: 70.6±0.45 to 68.9±0.81 (OPLU to Abs, updU) | No clear effect |
| Increase input size from 32 to 256 (ablation) | UGNN+MaxMin+updB · CIFAR10, Raw | accuracy: 70.4±0.52 -> 78.4±0.29 | Improved |
| Increase input size from 32 to 256 (ablation) | UGNN+MaxMin+updB · CIFAR10, Std.Norm | accuracy: 72.6±0.79 -> 76.5±0.35 | Improved |
| Use the UGNN lower bound instead of the compared 1-Lipschitz models’ lower bound | UGNN · CIFAR10, Raw | LB/MAP: ResNet9 .34±.063; LargeConvNet .46±.057; LipConvNet5 .58±.069; UGNN+MaxMin+updU .71±.088 | Improved |
| Use the UGNN lower bound instead of the compared 1-Lipschitz models’ lower bound | UGNN · CIFAR10, Std.Norm | LB/MAP: ResNet9 .26±.036; LargeConvNet .44±.027; LipConvNet5 .52±.031; UGNN+MaxMin+updB .96±.044 | Improved |
| Use UGNN+MaxMin+updB-256-raw rather than the 1-Lipschitz models | UGNN · CIFAR10 certifiable ε-robust classifications | accuracy | Improved |
Closest papers by meaning
- Robust One-Class Classification with Signed Distance Function using 1-Lipschitz Neural Networks (ICML 2023 · similarity 0.86)
- Second-Order Provable Defenses against Adversarial Attacks (ICML 2020 · similarity 0.86)
- Fast Adversarial Robustness Certification of Nearest Prototype Classifiers for Arbitrary Seminorms (NeurIPS 2020 · similarity 0.85)
- Asymmetric Certified Robustness via Feature-Convex Neural Networks (NeurIPS 2023 · similarity 0.85)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.83)
- Gradient Flow Provably Learns Robust Classifiers for Orthonormal GMMs (ICML 2025 · similarity 0.83)
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.
