Research map / Adversarial robustness and security / Certified and empirical robustness
TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization
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 TWINS with AT instead of AT | Robust ResNet50 fine-tuning · CIFAR10, CIFAR100, Caltech256, CUB200, Stanford Dogs | clean accuracy: +1.47; +1.24; +0.96; +2.35; +4.89 | Helped |
| Use TWINS with AT instead of AT | Robust ResNet50 fine-tuning · CIFAR10, CIFAR100, Caltech256, CUB200, Stanford Dogs | AA accuracy: +1.35; +3.26; -0.16; -0.49; +2.07 | Mixed |
| Use TWINS with TRADES instead of TRADES | Robust ResNet50 fine-tuning · CIFAR10, CIFAR100, Caltech256, CUB200, Stanford Dogs | AA accuracy: +1.40; +1.89; -1.62; +0.47; +1.31 | Mixed |
| Retain pre-trained BN means and standard deviations in the Frozen Net (ablation) | TWINS · Five downstream datasets | clean and robust accuracy | Helped |
| Add one-epoch target-statistics warmup (ablation) | TWINS-AT and TWINS-TRADES · Five downstream datasets | AA accuracy: +0.56; +2.25; +0.07; +0.75; +1.50 for TWINS-AT vs AT | Mixed |
| Increase weight decay (ablation) | TWINS-AT · CIFAR10 | AA accuracy: 49.11; 48.51; 49.8; 48.51; 10.0 | Helped |
| Increase pre-training attack bound from 1/255 to 8/255 (ablation) | TWINS-AT · CIFAR100 | Robust Acc.: 25.97 -> 28.62 | Helped |
| Use Frozen Net statistics initialized to (0,1) rather than pre-trained statistics (ablation) | TWINS-Init · Five downstream datasets | clean and robust accuracy | Hurt |
| Vary the TWINS loss weight (ablation) | TWINS-AT · CIFAR10 | AA: 49.11; 48.51; 49.8; 48.51; 10.0 | Mixed |
Closest papers by meaning
- Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning (CVPR 2020 · similarity 0.90)
- Learning Adversarially Robust Sparse Networks via Weight Reparameterization (AAAI 2023 · similarity 0.89)
- Adversarial Robustness against Multiple and Single $l_p$-Threat Models via Quick Fine-Tuning of Robust Classifiers (ICML 2022 · similarity 0.89)
- TWINFUZZ: Dual-Model Fuzzing for Robustness Generalization in Deep Learning (AAAI 2026 · similarity 0.88)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 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.