Research map / Adversarial robustness and security / Certified and empirical robustness
RVCL: Evaluating the Robustness of Contrastive Learning via Verification
Zekai Wang, Weiwei Liu
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 |
|---|---|---|---|
| increase contrastive adversarial training strength eps_train from 0 to 8.8/255 | SimCLR/RoCL ResNet-18 · CIFAR-10 probabilistic verification | ACR_CL (x10^-2): 0.832 -> 5.71 | Improved |
| increase eps_train from 0 to 8.8/255 | ResNet-18 · CIFAR-10 PGD robust accuracy | robust accuracy (%): 0.15 -> 22.8 | Improved |
| increase eps_train from 0 to 8.8/255 | ResNet-18 · CIFAR-10 AutoAttack robust accuracy | robust accuracy (%): 0 -> 19.7 | Improved |
| increase eps_train from 0 to 8.8/255 | ResNet-18 · CIFAR-10 clean accuracy | clean accuracy (%): 84.8 -> 75.5 | Hurt |
| increase eps_train from 0 to 8.8/255 | ResNet-50 · CIFAR-10 clean accuracy | clean accuracy (%): 88.2 -> 79.2 | Hurt |
| increase eps_train from 0 to 8.8/255 | ResNet-18 · CIFAR-100 PGD robust accuracy | robust accuracy (%): 0 -> 12.6 | Improved |
| increase eps_train from 0 to 8.8/255 | ResNet-18 · CIFAR-100 probabilistic verification | ACR_CL (x10^-2): 1.60 -> 5.86 | Improved |
| use AutoAttack instead of PGD as evaluation attack | deterministic verification models · Mnist/CIFAR-10 robust accuracy | robust accuracy: 5-7% drop | Hurt |
| use AutoAttack instead of PGD as evaluation attack | probabilistic verification models · CIFAR-10/100 robust accuracy | robust accuracy: 2-5% drop | Hurt |
| adversarially train the downstream layer (eps_trainLE=0.15 vs clean-trained) (ablation) | CL encoder with eps_train=0 on Mnist · Mnist robust accuracy | robust accuracy: 0% -> 87.2% (eps_test 0.15 with eps_trainLE 0.05 vs 0.15) | Improved |
| use CBC instead of CROWN verifier for computing ACR_CL (ablation) | RVCL · Mnist/CIFAR-10 ACR_CL | ACR_CL | No clear effect |
| use CBC instead of CROWN verifier (ablation) | RVCL · verification time | time per image: about 300 times slower | Hurt |
Closest papers by meaning
- Robustness Verification for Contrastive Learning (ICML 2022 · similarity 0.98)
- When does Contrastive Learning Preserve Adversarial Robustness from Pretraining to Finetuning? (NeurIPS 2021 · similarity 0.90)
- Decoupled Adversarial Contrastive Learning for Self-Supervised Adversarial Robustness (ECCV 2022 · similarity 0.89)
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.89)
- Investigating Why Contrastive Learning Benefits Robustness against Label Noise (ICML 2022 · similarity 0.89)
- Efficient Adversarial Contrastive Learning via Robustness-Aware Coreset Selection (NeurIPS 2023 · similarity 0.88)
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.
