AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

Double-Win Quant: Aggressively Winning Robustness of Quantized Deep Neural Networks via Random Precision Training and Inference

Yonggan Fu, Qixuan Yu, Meng Li, Vikas Chandra, Yingyan Lin

ICML 2021 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency · 9 citations (OpenAlex)

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 changedWhereMeasuredEffect
add RPI (random precision inference) to adversarially trained model (ablation)FGSM, FGSM-RS, PGD-7 on PreActResNet18/WideResNet32 · CIFAR-10 PGD-20robust accuracy: e.g. PGD-7 PreActResNet18 51.17 -> 57.09Improved
add RPT (random precision training with switchable BN) on top of RPI (ablation)PGD-7 / FGSM / FGSM-RS DWQ · CIFAR-10 PGD-20robust accuracy: PGD-7 PreActResNet18 57.09 -> 65.15Improved
add RPI+RPT on top of PGD-7 trainingPreActResNet18 CIFAR-100 · CIFAR-100 PGD-20robust accuracy: 27.97 -> 41.74Improved
add RPI+RPT on top of FGSM-RS on ImageNet ResNet-50 while reducing BitOPsResNet-50 ImageNet · ImageNet PGD-10robust accuracy: 30.28 -> 37.93Improved
add RPI+RPT on top of PGD-7 on MobileNetV2 (compact model)MobileNetV2 CIFAR-10 · CIFAR-10 PGD-20robust accuracy vs natural accuracy: robust 50.03 -> 57.72 but natural 80.40 -> 73.97Mixed
add RPI+RPT on top of PGD-7 under larger epsilonPreActResNet18/WideResNet32/MobileNetV2 CIFAR-10 · CIFAR-10 eps=12 and 16 PGD-20robust accuracy: gains of 11.66%-18.89% (eps=12) and 15.68%-23.26% (eps=16)Improved
add RPT with switchable BN vs CPT (cyclic precision training) on top of PGD-7 (ablation)PreActResNet18 CIFAR-10 · CIFAR-10 PGD-20robust accuracy: CPT-32 66.40 vs RPT 65.15 on PGDNo clear effect
replace RPT with CPT on FGSM and FGSM-RS training (ablation)PreActResNet18 CIFAR-10 · CIFAR-10 PGD-20robust accuracy: FGSM CPT-16 35.45 vs RPT 64.08Hurt
RPI with different training precisions (fixed-precision adversarial training) on PGD-7 networks (ablation)three PGD-7 networks · CIFAR-10 PGD-20natural and robust accuracy stability: within 1.12%No clear effect
RPI with smaller precision ranges and precisions close to training precision (no RPT) (ablation)PGD-7 + RPI · CIFAR-10 PGD-20robust accuracy: slightly favors smaller precision rangesMixed
add RPI+RPT on top of PGD-7 against stronger attacks (CW-Inf, Bandits, Auto-Attack)PreActResNet18/WideResNet32 CIFAR-10/100 · CW-Inf, Bandits, Auto-Attackrobust accuracy: +5.01% to +25.71% across attacksImproved
add RPI+RPT on top of PGD-7 comparison to Lin et al. 2019 reported resultsPGD-7 CIFAR-10 · CIFAR-10 PGD-20 eps=8 and 16robust accuracy: 14.6% and 22.5% higherImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy