Research map / Adversarial robustness and security / Certified and empirical robustness
Defensive Quantization: When Efficiency Meets Robustness
Ji Lin, Chuang Gan, Song Han
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 Lipschitz regularization to 5-bit activation quantization | Wide ResNet-28-10 · CIFAR-10 test, FGSM ε=8 | accuracy: 19.0 -> 51.8 | Improved |
| Add DQ to normal training | Wide ResNet-16-4 · SVHN, PGD ε=8 | accuracy: 05 -> 10 | Improved |
| Add DQ to normal training | Wide ResNet-28-10 · CIFAR-10, PGD ε=8 | accuracy: 01 -> 04 | Improved |
| Add DQ to R+FGSM adversarial training | Wide ResNet-16-4 · SVHN, PGD ε=8 | accuracy: 30 -> 47 | Improved |
| Add DQ to R+FGSM adversarial training | Wide ResNet-28-10 · CIFAR-10, PGD ε=8 | accuracy: 43 -> 50 | Improved |
| Add DQ to PGD adversarial training | Wide ResNet-28-10 · CIFAR-10, PGD ε=8 | accuracy: 44 -> 50 | Improved |
| Add DQ regularization to 4-bit ReLU1 quantization | ResNet · CIFAR-10, clean | accuracy: 93.49% -> 95.14% | Improved |
| Add DQ regularization to 4-bit ReLU6 quantization | ResNet · CIFAR-10, clean | accuracy: 94.41% -> 95.09% | Improved |
| Use Tanh-based rather than full-precision activation quantization | PGD-trained ResNet-18 · CIFAR-10 PGD white-box | accuracy: 50.51% -> 71.03% | Improved |
| Use Tanh-based rather than full-precision activation quantization | PGD-trained ResNet-18 · CIFAR-10 PGD black-box | accuracy: 65.58% -> 61.57% | Hurt |
| Use ReLU6 quantization rather than full precision | PGD-trained ResNet-18 · CIFAR-10 PGD black-box | accuracy: 65.58% -> 59.44% | Hurt |
| Replace adversarial FGSM training with R+FGSM training (ablation) | Wide ResNet · CIFAR-10, FGSM black-box ε=8 | accuracy: 51.65% -> 87.39% | Improved |
Closest papers by meaning
- Double-Win Quant: Aggressively Winning Robustness of Quantized Deep Neural Networks via Random Precision Training and Inference (ICML 2021 · similarity 0.91)
- Improved Gradient-Based Adversarial Attacks for Quantized Networks (AAAI 2022 · similarity 0.91)
- Qu-ANTI-zation: Exploiting Quantization Artifacts for Achieving Adversarial Outcomes (NeurIPS 2021 · similarity 0.90)
- Attacking Binarized Neural Networks (ICLR 2018 · similarity 0.89)
- Mind the Gap: A Practical Attack on GGUF Quantization (ICML 2025 · similarity 0.87)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.87)
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.
