AI Paper Atlas

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

Improving Generalization of Adversarial Training via Robust Critical Fine-Tuning

Kaijie Zhu, Xixu Hu, Jindong Wang, Xing Xie, Ge Yang

ICCV 2023 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training recipes · 45 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
fine-tune the lowest-MRC (non-robust-critical) module and interpolate weightsAdversarial training · CIFAR10/CIFAR100/Tiny-ImageNetStd accuracy: +1.21 / +2.02 / +1.87 avgImproved
apply RiFT to adversarially trained modelAdversarial training · CIFAR10-C/CIFAR100-C/Tiny-ImageNet-COOD accuracy: +1.49 / +1.81 / +1.75 avgImproved
apply RiFT to adversarially trained modelAdversarial training · CIFAR10/CIFAR100/Tiny-ImageNetAdv accuracy (PGD-10): +0.02 / +0.02 / +0.08 avgNo clear effect
apply RiFT to TRADES, MART, AWP, SCOREResNet18 · CIFAR10/CIFAR100Std accuracy: SCORE CIFAR100 54.83 -> 57.63Improved
fine-tune non-robust-critical module instead of all layers (ablation)RiFT · CIFAR10 (ResNet18)Adv accuracy (PGD-10): 52.66 -> 53.65Improved
fine-tune robust-critical module instead of non-robust-critical module (ablation)RiFT · CIFAR10 (ResNet18)Adv accuracy (PGD-10): 53.65 -> 52.48Hurt
fine-tune last layer only (linear probing) (ablation)RiFT · CIFAR10 (ResNet18)Adv accuracy (PGD-10): 53.65 -> 52.75Hurt
fine-tune more non-robust-critical modules (top 2, 3, 5) instead of top 1 (ablation)RiFT · CIFAR10 (ResNet18)Adv accuracy (PGD-10): 53.65 -> 52.47 (top 2)Hurt
apply RiFT to WRN34-10 on CIFAR10WRN34-10 · CIFAR10Std accuracy: 87.41 -> 87.89Improved

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