Research map / Adversarial robustness and security / Certified and empirical robustness
Metric Learning for Adversarial 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 |
|---|---|---|---|
| add adversarial triplet-loss metric regularization with semi-hard negative selection | LeNet and WRN classifiers · CIFAR-10, 20PGD (20 steps) | accuracy: 45.98% -> 50.03% | Helped |
| add adversarial triplet-loss metric regularization | LeNet · MNIST, 20PGD (100 steps) | accuracy: 94.93% -> 96.79% | Helped |
| add adversarial triplet-loss metric regularization | ResNet-50 · Tiny ImageNet, MIM (40 steps) | accuracy: 19.85% -> 20.69% | Helped |
| train with TLA rather than AT or ALP | MNIST and CIFAR-10 classifiers · unseen L0 and L2 attacks | accuracy: up to 1.10% and 13.1% | Helped |
| replace closest mini-batch negative sampling with random negative sampling (ablation) | TLA · CIFAR-10, 20PGD (20 steps) | accuracy: 50.03% -> 45.55% | Hurt |
| switch to a clean anchor and adversarial positive (ablation) | TLA · CIFAR-10, 20PGD (20 steps) | accuracy: 50.03% -> 49.15% | Hurt |
| increase negative mini-batch size from 1 to 500 (ablation) | TLA · CIFAR-10, FGSM (1 step) | accuracy: 55.41% -> 60.02% | Helped |
| increase negative mini-batch size from 500 to 2000 (ablation) | TLA · CIFAR-10, BIM (7 steps) | accuracy: 53.20% -> 52.52% | Hurt |
| remove the triplet-loss term by setting lambda to 0 (ablation) | TLA · MNIST, PGD (100 steps) | accuracy: 94.82% -> 96.96% | Hurt |
| vary triplet-loss weight across the tested values (ablation) | TLA · MNIST, PGD (100 steps) | accuracy: 94.82%; 96.31%; 96.96%; 96.57%; 96.72%; 96.26% | Mixed |
| vary triplet margin across the tested values (ablation) | TLA · MNIST, PGD (100 steps) | accuracy: 96.60%; 96.47%; 96.72%; 96.36%; 96.35% | Mixed |
| apply triplet loss at the higher penultimate representation rather than lower, middle, or logit layers (ablation) | TLA · MNIST, PGD (100 steps) | accuracy: 96.14%; 96.48%; 96.72%; 96.34% | Helped |
Closest papers by meaning
- Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning (Student Abstract) (AAAI 2023 · similarity 0.93)
- Provably Robust Metric Learning (NeurIPS 2020 · similarity 0.92)
- Enhancing Adversarial Robustness for Deep Metric Learning (CVPR 2022 · similarity 0.91)
- Measuring Neural Net Robustness with Constraints (NeurIPS 2016 · similarity 0.89)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (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.