Research map / Architectures and efficiency / RNNs and normalization
Robust Weight Signatures: Gaining Robustness as Easy as Patching Weights?
Ruisi Cai, Zhenyu Zhang, Zhangyang Wang
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 |
|---|---|---|---|
| add shallow-5-layer RWS (16-bit) to standard model | VGG-16 standard model · CIFAR-10, CIFAR-100, Tiny-ImageNet, ResNet-50 Tiny-ImageNet | averaged robust accuracy: +19.32 / +21.86 / +20.67 (VGG-16); +23.16 (ResNet-50) | Improved |
| patch with shallow RWS, then remove it to restore clean model | RWS patching · CIFAR-10 VGG-16 | clean accuracy TA: TA stays 92.59 with RWS taken off | No clear effect |
| restrict RWS to shallowest 5 layers instead of full model (ablation) | RWS patching · CIFAR-10 VGG-16 | RA (%): 84.86 vs 75.35 (full RWS) | Improved |
| 16-bit quantization of shallow RWS (ablation) | shallow RWS patch · CIFAR-10 VGG-16 | RA (%): 84.76 vs 84.86 (unquantized shallow) | No clear effect |
| 8-bit quantization of shallow RWS (ablation) | shallow RWS patch · CIFAR-10 VGG-16 | RA (%): 82.99 vs 84.76 (16-bit) | Hurt |
| quantize full model weights to 16 bit (ablation) | standard model · CIFAR-10 VGG-16 | RA (%): 58.12 vs 65.44 | Hurt |
| RWS storage compression (shallow 16-bit vs full model) | RWS patch · CIFAR-10 VGG-16 storage | N_param (MB): 79.9 MB vs 1177.6 MB (All Models) | Improved |
| transfer RWS from Tiny-ImageNet to ImageNet VGG-16 model | ImageNet standard model · ImageNet robust accuracy | robust accuracy (%): +6.39 (17.53 vs 11.14) | Improved |
| transfer RWS from CIFAR-10 to a model trained on another dataset (zero-shot) (ablation) | CIFAR-100 and Tiny-ImageNet standard models · cross-dataset RWS transfer | averaged robust accuracy: CIFAR-100 +18.89; Tiny-ImageNet +8.65 | Improved |
| data augmentation with matched FLOPS (CIFAR-10 budget) for ImageNet robustness | ImageNet VGG-16 · ImageNet robust accuracy | robust accuracy (%): -0.13 (11.01 vs 11.14) | Hurt |
Closest papers by meaning
- Improving robustness against common corruptions by covariate shift adaptation (NeurIPS 2020 · similarity 0.87)
- Benchmarking Low-Shot Robustness to Natural Distribution Shifts (ICCV 2023 · similarity 0.87)
- A Flat Minima Perspective on Understanding Augmentations and Model Robustness (AAAI 2026 · similarity 0.86)
- Foundation Model-oriented Robustness: Robust Image Model Evaluation with Pretrained Models (ICLR 2024 · similarity 0.85)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 2024 · similarity 0.84)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.84)
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.
