Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Contrastive Learning via Asymmetric InfoNCE
Qiying Yu, Jieming Lou, Xianyuan Zhan, Qizhang Li, Wangmeng Zuo, Yang Liu, Jingjing 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 |
|---|---|---|---|
| replace RoCL loss with inferior-positives loss L_IP | RoCL · CIFAR-10 Linear Probing | standard accuracy: 83.84 -> 87.63 | Improved |
| replace RoCL loss with L_IP+HN | RoCL · CIFAR-10 Linear Probing | robust accuracy: 38.98 -> 42.96 | Improved |
| replace RoCL loss with L_IP | RoCL · CIFAR-10 Adversarial Full Finetuning | robust accuracy: 50.54 -> 50.29 | No clear effect |
| replace RoCL loss with L_IP+HN | RoCL · CIFAR-100 Linear Probing | robust accuracy: 18.49 -> 22.54 | Improved |
| replace AdvCL loss with L_IP+HN | AdvCL · CIFAR-10 Linear Probing | robust accuracy: 51.00 -> 52.65 | Improved |
| replace AdvCL loss with L_IP+HN | AdvCL · CIFAR-10 Adversarial Linear Finetuning | robust accuracy: 52.38 -> 53.19 | No clear effect |
| replace AdvCL loss with L_HN | AdvCL · CIFAR-100 Adversarial Linear Finetuning | standard accuracy: 47.45 -> 47.32 | Hurt |
| add hard negatives L_HN to AdvCL, transfer CIFAR-100 to STL-10 | AdvCL · CIFAR-100 to STL-10 Adversarial Linear Finetuning | robust accuracy: 32.14 -> 33.62 | Improved |
| use L_IP+HN for AdvCL under AutoAttack | AdvCL · CIFAR-100 Linear Probing AutoAttack | robust accuracy: 19.57 -> 21.43 | Improved |
| use L_IP+HN instead of AdvCL under CIFAR-10 AutoAttack linear finetuning | AdvCL · CIFAR-10 Adversarial Linear Finetuning AutoAttack | robust accuracy: 43.93 -> 43.01 | Hurt |
| use fixed alpha instead of symmetric (alpha=0.5) (ablation) | CoreACL · CIFAR-10 | standard accuracy: 78.90 -> 81.29 | Improved |
| add adaptive alpha annealing (ablation) | CoreACL · CIFAR-10 | robust accuracy: 50.27 -> 51.37 | Improved |
Closest papers by meaning
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.86)
- When does Contrastive Learning Preserve Adversarial Robustness from Pretraining to Finetuning? (NeurIPS 2021 · similarity 0.86)
- Contrastive Learning with Adversarial Examples (NeurIPS 2020 · similarity 0.86)
- Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning (ICLR 2023 · similarity 0.85)
- Robustness Verification for Contrastive Learning (ICML 2022 · similarity 0.85)
- Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data (CVPR 2025 · similarity 0.84)
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.
