Research map / Adversarial robustness and security / Certified and empirical robustness
Rethinking the Effect of Data Augmentation in Adversarial 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 |
|---|---|---|---|
| add augmentation annealing (strong-to-weak schedule) to ACL (ablation) | ACL self-AT · CIFAR-10 | AA(%): +7.42 | Helped |
| add pseudo-label LP-AFT post-processing to ACL (ablation) | ACL self-AT · CIFAR-10 | AA(%): +5.62 | Helped |
| add both annealing and post-processing (DYNACL++) (ablation) | ACL self-AT · CIFAR-10 | AA(%): +8.84 | Helped |
| add augmentation annealing to ACL (ablation) | ACL self-AT · CIFAR-10 | SA(%): -1.91 | Hurt |
| add post-processing to ACL without annealing (ablation) | ACL self-AT · CIFAR-10 | SA(%): -1.92 | Hurt |
| use constant augmentation strength instead of annealing schedule (DYNACL++) (ablation) | DYNACL++ · CIFAR-10 | AA(%): -5.70 vs linear K=50 (40.76 vs 46.46) | Hurt |
| dynamic loss reweighting with lambda=2/3 vs none (lambda=0) (ablation) | DYNACL++ · CIFAR-10 | AA(%): +1.65 | Helped |
| add momentum encoder and stop-gradient to dynamic ACL (ablation) | DYNACL · CIFAR-10 | AA(%): +0.12 (44.92 to 45.04) | No clear effect |
| threshold augmentation strength at s=0.3 instead of annealing to 0 (ablation) | DYNACL++ · CIFAR-10 | AA(%): -0.57 | Hurt |
| DYNACL (augmentation annealing, no post-processing) vs prior self-AT on CIFAR-100 and STL-10 | self-AT · CIFAR-100 / STL-10 | AA(%): CIFAR-100 +3.57 over ACL (19.25 vs 15.68); STL-10 DYNACL++ +1.95 over best baseline | Helped |
| use medium static augmentation strength (s=0.5) for ACL | ACL self-AT · CIFAR-10 | robustness (PGD) | Mixed |
Closest papers by meaning
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 2024 · similarity 0.90)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · similarity 0.89)
- On the Duality Between Sharpness-Aware Minimization and Adversarial Training (ICML 2024 · similarity 0.87)
- Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR 2024 · similarity 0.87)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.87)
- Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data (CVPR 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.