Research map / Adversarial robustness and security / Certified and empirical robustness
Decoupled Adversarial Contrastive Learning for Self-Supervised Adversarial Robustness
Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Axi Niu, Jiu Feng, Chang D. Yoo, In So Kweon
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 |
|---|---|---|---|
| use two-stage decoupled training (DeACL) instead of single-stage AdvCL | adversarial contrastive learning · CIFAR10 (SLF) | AutoAttack accuracy (%): 42.57 -> 45.31 | Improved |
| use DeACL instead of AdvCL for pretraining cost | adversarial contrastive learning · CIFAR10 | total GPU-days: 12.60 -> 0.45 | Improved |
| use DeACL instead of AdvCL | adversarial contrastive learning · CIFAR100 (SLF) | standard accuracy (%): 48.34 -> 52.79 | Improved |
| raise stage-2 weight decay from 1e-6 to 5e-4 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 34.77 -> 45.31 | Improved |
| raise stage-2 weight decay from 5e-4 to 5e-3 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 24.80 | Hurt |
| use strong augmentation on both clean and adversarial examples at stage 2 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 6.93 | Hurt |
| use strong augmentation on clean examples only at stage 2 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 34.47 | Hurt |
| add projector head at stage 2 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 42.48 | Hurt |
| add contrastive collapse prevention at stage 2 (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 34.18 | Hurt |
| initialise student from pretrained stage-1 weights rather than scratch (ablation) | DeACL stage 2 · CIFAR10 (SLF) | AutoAttack accuracy (%): 44.63 -> 45.31 | No clear effect |
| swap SimCLR for MoCo v2 at stage 1 (ablation) | DeACL stage 1 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 46.29 | No clear effect |
| swap SimCLR for BarlowTwins at stage 1 (ablation) | DeACL stage 1 · CIFAR10 (SLF) | AutoAttack accuracy (%): 45.31 -> 41.31 | Hurt |
Closest papers by meaning
- Contrastive Self-Supervised Learning Leads to Higher Adversarial Susceptibility (AAAI 2023 · similarity 0.91)
- Adversarial Self-Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations From Self-Trained Negative Adversaries (CVPR 2021 · similarity 0.90)
- Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data (CVPR 2025 · similarity 0.88)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 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.
