Research map / Adversarial robustness and security / Certified and empirical robustness
Contrastive Self-Supervised Learning Leads to Higher Adversarial Susceptibility
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 |
|---|---|---|---|
| Train with self-supervised contrastive learning rather than supervised contrastive learning | ResNet50 · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 38.8 -> 26.8 | Hurt |
| Train with self-supervised contrastive learning rather than supervised contrastive learning | ResNet50 · CIFAR-100, FGSM-l∞ ε=8/255 | top-1 accuracy: 12.6 -> 9.40 | Hurt |
| Use SimCLR instead of supervised pretraining | ResNet50 · ImageNet, FGSM-l∞ ε=0.25/255 | top-1 accuracy: 38.20 -> 24.33 | Hurt |
| Use self-distillation without negative pairs rather than SimCLR | ResNet50 · ImageNet, FGSM-l∞ ε=0.25/255 | top-1 accuracy: 24.33 -> 39.40 | Helped |
| Use self-supervised contrastive learning rather than supervised training | R-(2+1)-D-18 · UCF101, FGSM-l∞ ε=1/255 | top-1 accuracy: 26.6 -> 10.8 | Hurt |
| Use self-supervised contrastive learning rather than supervised training | R-(2+1)-D-18 · HMDB51, FGSM-l∞ ε=1/255 | top-1 accuracy: 12.2 -> 5.7 | Hurt |
| Remove likely false-negative pairs using FNC | SimCLR · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 26.8 -> 28.6 | Helped |
| Remove likely false-negative pairs using IFND | SimCLR · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 26.8 -> 29.2 | Helped |
| Use a fixed threshold instead of an adaptive threshold (ablation) | Adaptive FNC · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 33.6 -> 30.1 | Hurt |
| Set the adaptive threshold schedule to ρ=0.9→0.5 (ablation) | Adaptive FNC · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 33.6 -> 31.3 | Hurt |
| Adaptively lower the threshold from ρ=0.99 to ρ=0.7 | SimCLR · CIFAR-10, FGSM-l∞ ε=8/255 | top-1 accuracy: 26.8 -> 33.6 | Helped |
| Adaptively lower the threshold from ρ=0.99 to ρ=0.7 | SimCLR · CIFAR-10, AutoAttack-l∞ ε=1/255 | top-1 accuracy: 13.7 -> 19.3 | Helped |
Closest papers by meaning
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.92)
- Adversarial Self-Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 2024 · similarity 0.88)
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 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.