Research map / Adversarial robustness and security / Certified and empirical robustness
Model Compression with Adversarial Robustness: A Unified Optimization Framework
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 |
|---|---|---|---|
| Jointly optimize adversarial training with the structured compression constraint rather than sequential adversarial pruning | ATMC-32bits · CIFAR-10 under WRM, pruning ratios 0.1, 0.05, 0.001 | ATA: 0.8271 -> 0.8331; 0.8190 -> 0.8289; 0.6952 -> 0.7311 | Helped |
| Jointly optimize adversarial training with structured compression rather than adversarial ℓ0 pruning | ATMC-32bits · CIFAR-10 under WRM, pruning ratios 0.1, 0.05, 0.001 | ATA: 0.8250 -> 0.8331; 0.8175 -> 0.8289; 0.7262 -> 0.7311 | Helped |
| Use 8-bit quantization instead of 32-bit precision | ATMC · CIFAR-10 under WRM, same pruning ratio | model size: only 1/4 model size | Helped |
| Use 8-bit quantization instead of 32-bit precision | ATMC · CIFAR-10 under WRM, pruning ratios 0.1, 0.05, 0.001 | ATA: 0.8331 -> 0.8112; 0.8289 -> 0.7996; 0.7311 -> 0.7144 | Hurt |
| Replace jointly learned non-uniform quantization with post-hoc uniform quantization (ablation) | ATMC-8bits · SVHN, pruning ratios 0.01 and 0.001 | ATA: 0.6954 -> 0.6855; 0.4028 -> 0.3533 | Hurt |
| Replace jointly learned non-uniform quantization with post-hoc uniform quantization (ablation) | ATMC-8bits · SVHN, pruning ratios 0.01 and 0.001 | TA: 0.9336 -> 0.9311; 0.7749 -> 0.7239 | Hurt |
| Use joint structured compression instead of adversarial ℓ0 pruning | ATMC-32bits · CIFAR-10 with PGD ε=8, pruning ratios 0.005, 0.003, 0.001 | ATA: 0.4813 -> 0.4875; 0.4559 -> 0.4645; 0.3532 -> 0.3608 | Helped |
| Use joint structured compression instead of adversarial pruning | ATMC-32bits · CIFAR-10 with PGD ε=8, pruning ratios 0.005, 0.003, 0.001 | ATA: 0.4569 -> 0.4875; 0.4247 -> 0.4645; 0.3398 -> 0.3608 | Helped |
| Use joint structured compression instead of adversarial ℓ0 pruning | ATMC-32bits · CIFAR-10 ResNet101 under WRM, pruning ratios 0.005, 0.001, 0.0008 | ATA: 0.6124 -> 0.6219; 0.4263 -> 0.4477; 0.4024 -> 0.4347 | Helped |
| Apply compression without adversarial defense | NAP · CIFAR-10 and CIFAR-100 | adversarial accuracy | Hurt |
| Train with WRM instead of PGD | ATMC · CIFAR-10 | adversarial accuracy | No clear effect |
| Increase PGD perturbation level to 8 for training and testing | ATMC · CIFAR-10 | adversarial accuracy | No clear effect |
Closest papers by meaning
- Rate Distortion For Model Compression:From Theory To Practice (ICML 2019 · similarity 0.89)
- Holistic Adversarially Robust Pruning (ICLR 2023 · similarity 0.88)
- Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks (NeurIPS 2025 · similarity 0.88)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.87)
- Out of Thin Air: Exploring Data-Free Adversarial Robustness Distillation (AAAI 2024 · similarity 0.85)
- Two is Better than One: Efficient Ensemble Defense for Robust and Compact Models (CVPR 2025 · similarity 0.85)
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.