Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning
Check 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 adversarial Selfie pretraining before end-to-end adversarial fine-tuning | CIFAR-10 classifier · CIFAR-10 | RA: 3.83% | Helped |
| Add adversarial Selfie pretraining before end-to-end adversarial fine-tuning | CIFAR-10 classifier · CIFAR-10 | TA: 1.3% | Helped |
| Add standard self-supervised pretraining before full adversarial fine-tuning | CIFAR-10 classifier · CIFAR-10 | RA: around 3% improvement | Helped |
| Add adversarial pretraining before adversarial fine-tuning | CIFAR-10 classifier · CIFAR-10 | Epochs: less than 50 epochs ... vs. 99 epochs | Helped |
| Use adversarial rather than standard full fine-tuning | Self-supervised pretrained CIFAR-10 classifier · CIFAR-10 | RA: 0.00 -> 50.42 | Helped |
| Use standard full fine-tuning on adversarially pretrained representations | Adversarially pretrained CIFAR-10 classifier · CIFAR-10 | RA: 51.05 -> 0.00 | Hurt |
| Use full rather than partial adversarial fine-tuning | Adversarially pretrained CIFAR-10 classifier · CIFAR-10 | RA | Mixed |
| Pretrain with Selfie rather than Jigsaw | Adversarially pretrained and full-AT fine-tuned classifier · CIFAR-10 | RA: around 2.5% | Helped |
| Add diversity-promoting regularization to an adversarial pretraining ensemble | AP + SE · CIFAR-10 | RA: at least 1.17% | Helped |
| Average predictions from three fine-tuned task-specific models | Selfie, Rotation, and Jigsaw adversarially pretrained models · CIFAR-10 | RA: 3.59% | Helped |
| Increase adversarial pretraining data from CIFAR-30K to CIFAR-150K (ablation) | Full adversarial fine-tuning pipeline · CIFAR-10 | RA: 0.97% | Helped |
| Replace adversarial pretraining with random-smoothing pretraining (ablation) | Selfie pretraining followed by adversarial fine-tuning · CIFAR-10 | RA | No clear effect |
Closest papers by meaning
- Robust Pre-Training by Adversarial Contrastive Learning (NeurIPS 2020 · similarity 0.92)
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.92)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.90)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · similarity 0.90)
- Long-tailed Adversarial Training with Self-Distillation (ICLR 2025 · similarity 0.89)
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.