Research map / Adversarial robustness and security / Certified and empirical robustness
Contrastive Learning with Adversarial Examples
Chih-Hui Ho, Nuno Nvasconcelos
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 |
|---|---|---|---|
| Add adversarial augmentation and adversarial training to Plain contrastive learning | Plain · CIFAR10, CIFAR100, tinyImageNet | downstream classification accuracy: 82.78$\pm$0.20 -> 83.09$\pm$0.19; 54.73$\pm$0.20 -> 55.28$\pm$0.12; 79.65$\pm$0.43 -> 79.94$\pm$0.28; 51.82$\pm$0.46 -> | Improved |
| Add adversarial augmentation and adversarial training to UEL | UEL · CIFAR10, CIFAR100, tinyImageNet | downstream classification accuracy: 83.63$\pm$0.14 -> 84$\pm$0.15; 55.23$\pm$0.28 -> 55.96$\pm$0.06; 80.63$\pm$0.18 -> 80.94$\pm$0.13; 52.99$\pm$0.25 -> 54. | Improved |
| Add adversarial augmentation and adversarial training to SimCLR | SimCLR · CIFAR10, CIFAR100, tinyImageNet | downstream classification accuracy: 75.92$\pm$0.26 -> 76.45$\pm$0.32; 34.94$\pm$0.25 -> 38.89$\pm$0.25; 83.27$\pm$0.17 -> 83.32$\pm$0.26; 53.79$\pm$0.21 -> | Improved |
| Use ε=0.07 rather than ε=0.03 (ablation) | CLAE · CIFAR10, CIFAR100, tinyImageNet | downstream classification accuracy: 83.09$\pm$0.19 -> 83.04$\pm$0.18; 55.28$\pm$0.12 -> 54.96$\pm$0.12; 79.94$\pm$0.28 -> 79.85$\pm$0.16; 52.04$\pm$0.32 -> | Mixed |
| Increase batch size (ablation) | SimCLR and SimCLR CLAE · CIFAR100 | downstream classification accuracy | Improved |
| Use adversarial training at batch size 64 rather than the baseline at batch size 256 (ablation) | SimCLR CLAE vs SimCLR · CIFAR100 | downstream classification accuracy: 53.79\% -> 54\% | Improved |
| Vary embedding dimension (ablation) | SimCLR and SimCLR CLAE · CIFAR100 | downstream classification accuracy | No clear effect |
| Use a larger ResNet architecture (ablation) | SimCLR CLAE · CIFAR100 | downstream classification accuracy improvement: 5\% | Improved |
| Increase the adversarial loss weight α from 0 to around 0.8 (ablation) | SimCLR CLAE · CIFAR100 | downstream classification accuracy | Improved |
| Increase α beyond 0.2, up to 2 (ablation) | SimCLR CLAE · CIFAR100 | downstream classification accuracy | No clear effect |
| Increase attack strength ε across the tested range (ablation) | SimCLR CLAE · CIFAR100 | downstream classification accuracy | No clear effect |
| Train for more pretext epochs (ablation) | SimCLR and SimCLR CLAE · CIFAR100 | downstream classification accuracy: 300 epochs vs baseline 400 epochs; 400 epochs vs baseline 500 epochs | Improved |
Closest papers by meaning
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.88)
- Decoupled Adversarial Contrastive Learning for Self-Supervised Adversarial Robustness (ECCV 2022 · similarity 0.88)
- CaCo: Both Positive and Negative Samples are Directly Learnable via Cooperative-Adversarial Contrastive Learning (TPAMI 2023 · similarity 0.88)
- RVCL: Evaluating the Robustness of Contrastive Learning via Verification (JMLR 2023 · similarity 0.88)
- AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations From Self-Trained Negative Adversaries (CVPR 2021 · similarity 0.88)
- Contrastive Learning with Synthetic Positives (ECCV 2024 · similarity 0.88)
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.
