Research map / Adversarial robustness and security / Certified and empirical robustness
Integer-Arithmetic-Only Certified Robustness for Quantized Neural Networks
Haowen Lin, Jian Lou, Li Xiong, Cyrus Shahabi
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 |
|---|---|---|---|
| quantize to int8 with integer randomized smoothing instead of float network | randomized smoothing, ResNet-50 · Caltech-101 sigma=1.00, radius 2.0 | certified accuracy: 0.41 -> 0.52 | Improved |
| use IntRS-quant instead of FloatRS-quant (rounded float noise on quantized net) | quantized ResNet-20 · CIFAR-10 sigma=1.00, radius 0.25 | certified accuracy: 0.19 -> 0.35 | Improved |
| use IntRS-quant instead of FloatRS-quant | quantized ResNet-50 · Caltech-101 sigma=1.00, radius 0.25 | certified accuracy: 0.20 -> 0.56 | Improved |
| quantize with IntRS versus full-precision float smoothing at low sigma | ResNet-20 · CIFAR-10 sigma=0.50, radius 0.25 | certified accuracy: 0.54 -> 0.52 | Hurt |
| quantize with IntRS versus full-precision float smoothing at sigma=1.00 | ResNet-20 · CIFAR-10 sigma=1.00, radius 0.25 | certified accuracy: 0.39 -> 0.35 | Hurt |
| int8 integer-only inference instead of float inference | IntRS · CIFAR-10 certification time | seconds per image: 194.4 -> 55.8 | Improved |
| int8 integer-only inference instead of float inference | IntRS · Caltech-101 certification time | seconds per image: 159.9 -> 29.5 | Improved |
| increase noise sigma used in training and certification (ablation) | IntRS-quant · CIFAR-10 sigma 0.12 -> 1.00 | certified accuracy at radius 0.25: 0.59 -> 0.35 | Hurt |
| certify with sigma=0.25 on model trained with sigma=0.12 (mismatch) (ablation) | IntRS-quant · CIFAR-10 | certified accuracy at radius 0.25: 0.59 -> 0.36 | Hurt |
| IntRS under train/certify sigma mismatch versus Float-fp | IntRS-quant · CIFAR-10 trained sigma=0.12, certified sigma=0.25 | certified accuracy at radius 0.25: 0.19 -> 0.36 | Improved |
| increase Monte-Carlo samples n from 100 to 10000 (ablation) | IntRS-quant · CIFAR-10 sigma=0.25 prediction | abstain rate: 0.15 -> 0.02 | Improved |
| increase Monte-Carlo samples n from 100 to 10000 (ablation) | IntRS-quant · CIFAR-10 sigma=0.25 prediction | correct rate: 0.73 -> 0.80 | Improved |
Closest papers by meaning
- Black-Box Certification with Randomized Smoothing: A Functional Optimization Based Framework (NeurIPS 2020 · similarity 0.89)
- Certified Adversarial Robustness via Randomized $\alpha$-Smoothing for Regression Models (NeurIPS 2024 · similarity 0.89)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.88)
- Incremental Randomized Smoothing Certification (ICLR 2024 · similarity 0.88)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.87)
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (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.
