Research map / Adversarial robustness and security / Certified and empirical robustness
When does Contrastive Learning Preserve Adversarial Robustness from Pretraining to Finetuning?
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 an adversarial view generated from the original image, alongside the two augmented views (ablation) | AdvCL contrastive pretraining · CIFAR-10 SLF | RA: 42.12 -> 50.19 | Helped |
| add a high-frequency image view (ablation) | AdvCL contrastive pretraining · CIFAR-10 SLF | RA: 50.19 -> 50.45 | Helped |
| use a low-frequency view rather than a high-frequency view (ablation) | AdvCL contrastive pretraining · CIFAR-10 SLF | RA: 50.45 -> 49.51 | Hurt |
| add both low- and high-frequency views (ablation) | AdvCL contrastive pretraining · CIFAR-10 SLF | RA: 50.45 -> 50.03 | Hurt |
| add ClusterFit pseudo-label adversarial classification stimulus (ablation) | AdvCL · CIFAR-10 SLF | RA: 48.89 -> 50.45 | Helped |
| ensemble pseudo-label cluster counts instead of using a single cluster count (ablation) | AdvCL with ImageNet SimCLR ClusterFit · CIFAR-10 SLF | RA: 50.12 -> 50.45 | Helped |
| use standard linear finetuning instead of adversarial linear finetuning (ablation) | AdvCL · CIFAR-10 | RA: 52.01 -> 50.45 | Hurt |
| use AdvCL pretraining rather than ACL | ResNet-18 · CIFAR-10 SLF | RA: 42.87 -> 50.45 | Helped |
| use AdvCL pretraining rather than ACL | ResNet-18 · CIFAR-100 SLF | RA: 20.97 -> 27.67 | Helped |
| use AdvCL pretraining rather than ACL | ResNet-18 · CIFAR-100 → STL-10 SLF | AA: 21.75 -> 20.86 | Hurt |
| use AdvCL pretraining rather than ACL | ResNet-18 · CIFAR-100 → STL-10 SLF | SA: 45.91 -> 50.71 | Helped |
| use AdvCL pretraining rather than ACL | ResNet-18 · SVHN | RA: 39.34 -> 45.15 | Helped |
Closest papers by meaning
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.93)
- On Transfer of Adversarial Robustness from Pretraining to Downstream Tasks (NeurIPS 2023 · similarity 0.91)
- Robustness Verification for Contrastive Learning (ICML 2022 · similarity 0.91)
- Investigating Why Contrastive Learning Benefits Robustness against Label Noise (ICML 2022 · similarity 0.90)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.89)
- 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.