Research map / Architectures and efficiency / Neural architecture search
TRNAS: A Training-Free Robust Neural Architecture Search
Yeming Yang, Qingling Zhu, Jianping Luo, Ka-Chun Wong, Qiuzhen Lin, Jianqiang Li
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 |
|---|---|---|---|
| Use TRNAS searched architecture instead of ZCPRob under white-box attacks | TRNAS · CIFAR-10 | AutoAttack accuracy (%): 49.97 -> 51.76 | Improved |
| Use TRNAS instead of ZCPRob on CIFAR-100 | TRNAS · CIFAR-100 | AutoAttack accuracy (%): 25.59 -> 27.32 | Improved |
| Transfer TRNAS architecture to ImageNet | TRNAS · ImageNet | AutoAttack accuracy (%): 8.81 -> 9.08 (vs SWAP) | Improved |
| Use proxy-based search without any training instead of training-based robust NAS | TRNAS · CIFAR-10 search cost | GPU days: 0.02 vs 0.2 (CRoZe) | Improved |
| Replace robust NTK proxy in ZCPRob baseline with R-Score (ablation) | ZCPRob baseline · CIFAR-10 (Fig. 8) | robust accuracy | Mixed |
| Add multi-objective selection (MOS) on top of R-Score (ablation) | TRNAS · CIFAR-10 (Fig. 8) | clean and robust accuracy | Improved |
| Use R-Score instead of other zero-cost proxies (ablation) | TRNAS · RobustBench | Spearman correlation with robustness | Improved |
| Use gradient-consistency proxy (CRoZe) compared to R-Score | CRoZe · Proxy evaluation, 2000 architectures | effective rate: 79.30 % vs 100.0 % | Hurt |
| Use classic clean-accuracy zero-cost proxies for robustness | Zero-cost proxies · RobustBench | correlation with adversarial robustness | Hurt |
| Use TRNAS searched architecture as black-box target and as source | TRNAS · CIFAR-10 transfer attacks | accuracy under transfer attack: 65.57 target TRNAS vs 62.93 as source | Improved |
Closest papers by meaning
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.94)
- Towards Accurate and Robust Architectures via Neural Architecture Search (CVPR 2024 · similarity 0.93)
- Generalizable Lightweight Proxy for Robust NAS against Diverse Perturbations (NeurIPS 2023 · similarity 0.92)
- DSRNA: Differentiable Search of Robust Neural Architectures (CVPR 2021 · similarity 0.92)
- RNAS-CL: Robust Neural Architecture Search by Cross-Layer Knowledge Distillation (IJCV 2024 · similarity 0.91)
- Robustifying and Boosting Training-Free Neural Architecture Search (ICLR 2024 · 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.
