Research map / Adversarial robustness and security / Certified and empirical robustness
Quantization-Aware Interval Bound Propagation for Training Certifiably Robust Quantized Neural Networks
Mathias Lechner, Đorđe Žikelić, Krishnendu Chatterjee, Thomas A. Henzinger, Daniela Rus
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 |
|---|---|---|---|
| Train quantized networks with QA-IBP rather than prior QF_BV SMT-certified models | QNNs · MNIST and Fashion-MNIST at ε=0, 1, and 4 | certified robust accuracy: 99.2% vs 97.1%; 98.8% vs 92%; 95.6% vs 53%; 86.3% vs 85.6%; 80.0% vs 76%; 59.8% vs 36% | Improved |
| Use QA-IBP instead of the strongest prior QF_BV SMT result | QNNs · MNIST ε=1 | certified robust accuracy: 99% vs 98.8% | Hurt |
| Add branching into subdomains to IBP bounds plus PGD (ablation) | QA-IBP verifier · MNIST ε=4 | undetermined samples: 3.51% -> 3.47% | Improved |
| Add branching into subdomains to IBP bounds plus PGD (ablation) | QA-IBP verifier · Fashion-MNIST ε=1 and ε=4 | undetermined samples: 3.96% -> 3.93%; 19.23% -> 18.51% | Improved |
| Add branching into subdomains to IBP bounds plus PGD (ablation) | QA-IBP verifier · MNIST ε=1 | undetermined samples: 0.35% -> 0.35% | No clear effect |
| Reduce weight decay from 1·10^-4 to 1·10^-5 (ablation) | QA-IBP CIFAR-10 model · CIFAR-10 ε=0 versus ε=1 and ε=4 | certified robust accuracy: 30.0% -> 83.1%; 28.4% -> 0%; 22.4% -> 0% | Mixed |
| Use weight decay 5·10^-5 instead of 1·10^-4 (ablation) | QA-IBP CIFAR-10 model · CIFAR-10 ε=0 and ε=1 versus ε=4 | certified robust accuracy: 30.0% -> 39.6%; 28.4% -> 34.1%; 22.4% -> 20.6% | Mixed |
Closest papers by meaning
- Defensive Quantization: When Efficiency Meets Robustness (ICLR 2019 · similarity 0.89)
- Integer-Arithmetic-Only Certified Robustness for Quantized Neural Networks (ICCV 2021 · similarity 0.89)
- Improved Gradient-Based Adversarial Attacks for Quantized Networks (AAAI 2022 · similarity 0.89)
- Understanding Certified Training with Interval Bound Propagation (ICLR 2024 · similarity 0.88)
- Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · similarity 0.86)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.86)
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.
