Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarially Robust Few-Shot Learning: A Meta-Learning Approach
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 |
|---|---|---|---|
| adversarial querying (attack query data during meta-training) instead of adversarial-training transfer | R2-D2 · 5-shot Mini-ImageNet | robust accuracy: 25.33 -> 31.52 | Helped |
| adversarial querying instead of ADML | MAML-family · 5-shot Mini-ImageNet | robust accuracy: 18.49 -> 31.52 | Helped |
| adversarial querying vs natural training | R2-D2 · 5-shot Mini-ImageNet | robust accuracy: 0.00 -> 31.52 | Helped |
| adversarial querying vs natural training | R2-D2 · 5-shot Mini-ImageNet | natural accuracy: 73.02 -> 57.87 | Hurt |
| also perturb support data during meta-training (ablation) | MAML AQ · 5-shot Mini-ImageNet | robust accuracy: 23.07 -> 22.55 | No clear effect |
| also perturb support data during meta-training (ablation) | MAML AQ · 5-shot Omniglot | robust accuracy: 95.85 -> 77.73 | Hurt |
| fine-tune only last FC layer (ablation) | MAML AQ · 5-shot Mini-ImageNet | robust accuracy: 23.07 -> 25.15 | Helped |
| fine-tune only last FC layer (ablation) | MAML AQ · 5-shot Mini-ImageNet | natural accuracy: 33.45 -> 40.06 | Helped |
| add feature denoising blocks (ablation) | R2-D2 AQ · 5-shot Mini-ImageNet | robust accuracy: 31.52 -> 31.14 | No clear effect |
| use TRADES loss in query step (meta-TRADES, 1/lambda=3) (ablation) | R2-D2 AQ · 5-shot Mini-ImageNet | robust accuracy: 31.52 -> 32.30 | Mixed |
| replace ridge-regression head with SVM head (ablation) | R2-D2 AQ · 5-shot Mini-ImageNet | robust accuracy: 31.52 -> 28.08 | Hurt |
| add DefenseGAN pre-processing to natural R2-D2 instead of AQ (ablation) | R2-D2 · 5-shot CIFAR-FS | robust accuracy: 44.80 -> 28.05 | Hurt |
Closest papers by meaning
- MetaGAN: An Adversarial Approach to Few-Shot Learning (NeurIPS 2018 · similarity 0.90)
- Adversarially Robust Few-shot Learning via Parameter Co-distillation of Similarity and Class Concept Learners (CVPR 2024 · similarity 0.90)
- Improving Adversarially Robust Few-Shot Image Classification With Generalizable Representations (CVPR 2022 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- On Fast Adversarial Robustness Adaptation in Model-Agnostic Meta-Learning (ICLR 2021 · similarity 0.88)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 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.