Research map / Adversarial robustness and security / Certified and empirical robustness
Provably Robust Metric Learning
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 |
|---|---|---|---|
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | 1-NN metric learning · MNIST | certified robust error at ℓ2 radius 0.500: 0.112 -> 0.089 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | 1-NN metric learning · Fashion-MNIST | certified robust error at ℓ2 radius 0.500: 0.381 -> 0.348 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | 1-NN metric learning · Splice | certified robust error at ℓ2 radius 0.200: 0.404 -> 0.370 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | K-NN metric learning · MNIST | certified robust error at ℓ2 radius 0.500: 0.175 -> 0.101 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | K-NN metric learning · MNIST | empirical robust error at ℓ2 radius 0.500: 0.063 -> 0.055 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | K-NN metric learning · Fashion-MNIST | certified robust error at ℓ2 radius 0.500: 0.420 -> 0.371 | Helped |
| Learn the Mahalanobis metric by optimizing a lower bound on adversarial perturbation | K-NN metric learning · Splice | certified robust error at ℓ2 radius 0.200: 0.415 -> 0.345 | Helped |
| Use neighborhood size 10 for sampling positive and negative instances | ARML · Splice and Satimage | certified and empirical robust errors | No clear effect |
| Use NCA or LMNN metric learning instead of Euclidean distance | 1-NN and K-NN · Most evaluated datasets | certified robust error | Hurt |
| Use ARML instead of previous metric learning methods | 1-NN metric learning · Evaluated datasets | clean error: best on 4/6 | Mixed |
Closest papers by meaning
- Metric Learning for Adversarial Robustness (NeurIPS 2019 · similarity 0.92)
- Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning (Student Abstract) (AAAI 2023 · similarity 0.89)
- Enhancing Adversarial Robustness for Deep Metric Learning (CVPR 2022 · similarity 0.89)
- Unsupervised Metric Learning with Synthetic Examples (AAAI 2020 · similarity 0.88)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.85)
- Collapse-Aware Triplet Decoupling for Adversarially Robust Image Retrieval (ICML 2024 · similarity 0.85)
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.