Research map / Adversarial robustness and security / Certified and empirical robustness
Empirical Study of the Topology and Geometry of Deep 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 decision-boundary curvature asymmetry to detect small adversarial perturbations | ImageNet classifiers, including GoogLeNet · ImageNet | adversarial detection accuracy on clean samples: more than 95% | Helped |
| Use decision-boundary curvature asymmetry to recover the original label of perturbed samples | GoogLeNet · ImageNet | recovered-label accuracy: 92% | Helped |
| Use decision-boundary curvature asymmetry to recover the original label of perturbed samples | CaffeNet · ImageNet | recovered-label accuracy: 88% | Helped |
| Use decision-boundary curvature asymmetry to recover the original label of perturbed samples | VGG-19 · ImageNet | recovered-label accuracy: 74% | Helped |
| Use the shared highly curved directions as the perturbation subspace S | LeNet on CIFAR-10 · CIFAR-10 noisy validation set | misclassification rate under noise | Mixed |
| Use the sign of average boundary curvature to distinguish clean from adversarial samples | LeNet · CIFAR-10 test | curvature-sign classification rate: 97%; 96% | Helped |
| Use the sign of average boundary curvature to distinguish clean from adversarial samples | NiN · CIFAR-10 test | curvature-sign classification rate: 94%; 93% | Helped |
| Construct recursively adjusted paths between points with the same predicted label | CaffeNet · ImageNet validation | path existence within the classification region | Helped |
| Use recursively adjusted paths rather than requiring the straight segment to remain in the region | CaffeNet · ImageNet validation | path deviation D(p): 1 + 10−4 | Helped |
| Scale adversarial perturbations by a larger factor | Curvature-based detection on GoogLeNet · ImageNet | adversarial detection accuracy | Hurt |
Closest papers by meaning
- Robustness of Classifiers to Universal Perturbations: A Geometric Perspective (ICLR 2018 · similarity 0.91)
- Hold me tight! Influence of discriminative features on deep network boundaries (NeurIPS 2020 · similarity 0.89)
- Data Representations' Study of Latent Image Manifolds (ICML 2023 · similarity 0.89)
- Deep Networks and the Multiple Manifold Problem (ICLR 2021 · similarity 0.89)
- Geometric Inductive Biases of Deep Networks: The Role of Data and Architecture (ICLR 2025 · similarity 0.88)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · similarity 0.87)
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.