Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Self-Supervised Contrastive Learning
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 |
|---|---|---|---|
| use contrastive loss to generate instance-wise attacks instead of cosine similarity (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 9.30 -> 40.27 | Helped |
| use contrastive loss instead of Manhattan distance for attack (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 21.05 -> 40.27 | Helped |
| use contrastive loss instead of MSE for attack (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 40.12 -> 40.27 | No clear effect |
| generate attack on t'(x) instead of original x (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 36.6 -> 40.27 | Mixed |
| regularization with t'(x) instead of t(x) (ablation) | RoCL · CIFAR-10 PGD linf | accuracy: 29.97 -> 36.71 | Helped |
| add robust linear evaluation (class-adversarial linear classifier) (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 40.27 -> 47.69 | Helped |
| add robust linear evaluation (ablation) | RoCL · CIFAR-10 PGD l1 eps=12 (unseen) | accuracy: 76.17 -> 75.05 | Hurt |
| use RoCL self-supervised adversarial training instead of SimCLR | contrastive learning · CIFAR-10 PGD linf eps=8/255 | accuracy: 0.63 -> 40.27 | Helped |
| RoCL vs supervised AT on unseen l1 attack | RoCL · CIFAR-10 PGD l1 eps=12 | accuracy: 55.74 -> 76.17 | Helped |
| RoCL vs TRADES on seen linf eps=8/255 | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 48.01 vs 40.27 | Hurt |
| add transformation smoothed classifier | RoCL · CIFAR-10 clean | accuracy: 83.71 -> 84.11 | Helped |
| increase PGD attack steps from 20 to 100 (ablation) | RoCL · CIFAR-10 PGD linf eps=8/255 | accuracy: 40.27 -> 39.74 | No clear effect |
Closest papers by meaning
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.92)
- Contrastive Self-Supervised Learning Leads to Higher Adversarial Susceptibility (AAAI 2023 · similarity 0.91)
- Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data (CVPR 2025 · similarity 0.90)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 2024 · 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.