Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness vs. Model Compression, or Both?
Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, Xue Lin
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 |
|---|---|---|---|
| prune a w=4 model down to w=1 with concurrent adversarial training instead of training w=1 adversarially from scratch | LeNet · MNIST | natural / adversarial test accuracy: 11.35/11.35 -> 96.22/89.41 | Improved |
| start from w=8 instead of w=4 before pruning to w=1 | LeNet concurrent pruning · MNIST | natural / adversarial test accuracy: 96.22/89.41 -> 97.31/92.16 | Improved |
| start from very large w=16 rather than w=8 when pruning to w=1 | LeNet concurrent pruning · MNIST | adversarial test accuracy: 92.16 -> 87.79 | Hurt |
| inherit initialization from larger adversarially trained model (lottery ticket) without pruning, w=4 to 2 (ablation) | LeNet · MNIST | natural / adversarial test accuracy: 97.68/91.77 -> 11.35/11.35 | Hurt |
| inherited initialization from w=8 for w=2 model, no pruning (ablation) | LeNet · MNIST | natural / adversarial test accuracy: 97.36/90.19 | Mixed |
| adversarial training of a small w=1 model instead of natural training | LeNet · MNIST | natural test accuracy: 98.25 -> 11.35 | Hurt |
| adversarial training on CIFAR10 LeNet w=1 instead of natural training | LeNet · CIFAR10 | natural / adversarial test accuracy: 74.84/0.01 -> 10.00/10.00 | Hurt |
| prune ResNet w=16 to w=4 versus training w=4 adversarially from scratch | ResNet · CIFAR10 | natural / adversarial test accuracy: 77.35/44.99 -> 78.85/47.19 | Improved |
| irregular pruning instead of filter pruning (ablation) | VGG-16 and ResNet-18 · CIFAR10 | natural and adversarial accuracy | Improved |
| column pruning instead of filter pruning (ablation) | VGG-16 and ResNet-18 · CIFAR10 | natural and adversarial accuracy | Improved |
| choose different initializations (kaiming normal, uniform) with Adam for adversarial training from scratch at w=1 (ablation) | LeNet · MNIST | training success: 1 out of 10 trials succeeds | No clear effect |
| apply concurrent pruned models against transfer attack from larger baseline models | pruned LeNet · MNIST | adversarial test accuracy: 89.41 -> 93.55 | Improved |
Closest papers by meaning
- Learning Extremely Lightweight and Robust Model with Differentiable Constraints on Sparsity and Condition Number (ECCV 2022 · similarity 0.90)
- l∞-Robustness and Beyond: Unleashing Efficient Adversarial Training (ECCV 2022 · similarity 0.89)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.89)
- Holistic Adversarially Robust Pruning (ICLR 2023 · similarity 0.89)
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.89)
- 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.
