Research map / Adversarial robustness and security / Certified and empirical robustness
Data-Efficient Double-Win Lottery Tickets from Robust Pre-training
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 |
|---|---|---|---|
| use AT rather than standard or fast adversarial pre-training | ResNet-50 lottery tickets transferred to CIFAR-10, CIFAR-100, and SVHN · CIFAR-10, CIFAR-100, and SVHN | maximum matching sparsity: 83.22% ~ 96.48% vs. 20.00% ~ 89.26% and 0.00% ~ 89.26% | Helped |
| use robust pre-training instead of standard pre-training | data-limited transfer tickets · CIFAR-10, CIFAR-100, and SVHN | standard accuracy: up to 0.75%, 22.53% | Helped |
| use robust pre-training instead of standard pre-training | data-limited transfer tickets · CIFAR-10, CIFAR-100, and SVHN | robust accuracy: up to 0.79%, 8.97% | Helped |
| use FAT rather than AT pre-training at middle sparsity | data-limited transfer tickets · CIFAR-10, CIFAR-100, and SVHN | standard accuracy: 30.97%, 2.42%, 10.05% | Helped |
| use FAT rather than AT pre-training at middle sparsity | data-limited transfer tickets · CIFAR-10, CIFAR-100, and SVHN | robust accuracy: 8.36%, 0.70%, 18.49% | Helped |
| increase sparsity beyond the middle range | FAT- and AT-pre-trained tickets · CIFAR-10, CIFAR-100, and SVHN data-limited transfer | standard and robust accuracy: 8.98% and 9.62% | Mixed |
| use AT pre-training rather than FAT or standard pre-training | CIFAR-10 downstream transfer · CIFAR-10 | transfer attack accuracy: 55.54 -> 62.79 | Helped |
| use AT pre-training rather than FAT or standard pre-training | CIFAR-10 downstream transfer · CIFAR-10 | robust accuracy: 47.11 -> 56.00 | Helped |
| use adversarial rather than standard re-training during IMP (ablation) | AT-pre-trained ResNet-50 tickets · CIFAR-10 and CIFAR-100 | maximum matching sparsity: 91.41%, 83.22% vs. 91.41%, 83.22% | No clear effect |
| replace IMP re-training with one-shot magnitude pruning (ablation) | AT-pre-trained ResNet-50 tickets · CIFAR-10 and CIFAR-100 | maximum matching sparsity: 89.26%, 79.03% vs. 91.41%, 83.22% | Hurt |
| replace IMP with random pruning (ablation) | AT-pre-trained ResNet-50 tickets · CIFAR-10 and CIFAR-100 | maximum matching sparsity: 20%, 0% vs. 91.41%, 83.22% | Hurt |
| use AT-pre-trained tickets at higher sparsity on SVHN with limited data | robust-pretrained transfer tickets · SVHN, 10% and 1% training data | standard and robust accuracy: 83.22% and 97.19% | Hurt |
Closest papers by meaning
- Winning Lottery Tickets in Deep Generative Models (AAAI 2021 · similarity 0.91)
- Dual Lottery Ticket Hypothesis (ICLR 2022 · similarity 0.91)
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.90)
- The Lottery Tickets Hypothesis for Supervised and Self-Supervised Pre-Training in Computer Vision Models (CVPR 2021 · similarity 0.90)
- Lottery Pools: Winning More by Interpolating Tickets without Increasing Training or Inference Cost (AAAI 2023 · similarity 0.90)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · 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.