AI Paper Atlas

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

Adversarial Parameter Attack on Deep Neural Networks

Lijia Yu, Yihan Wang, Xiao-Shan Gao

ICML 2023 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training refinements · 1 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
Random L_inf parameter perturbations (no adversarial optimization) (ablation)VGG19 · CIFAR-10adversarial rate: AR 0.13 at ratio 0.02 to 0.24 at 0.10 with accuracy keptNo clear effect
Adversarial parameter optimization (L_inf attack) instead of random perturbation (ablation)VGG19 · CIFAR-10adversarial accuracy: AA drops from 45% to 8% at ratio 0.10 with AC 77%Improved
Increase L_inf perturbation ratio from 0.02 to 0.10 (ablation)VGG19 adversarial parameter attack · CIFAR-10adversarial rate: AR rises from 0.15 to 0.79Improved
Increase network depth (more layers at same width) (ablation)VGG19 adversarial parameter attack · CIFAR-10adversarial rate at gamma=0.04: AR 0.32 for VGG19 to 0.53 for 32 extra layersImproved
Increase network width (1.25x to 2.5x) (ablation)VGG19 adversarial parameter attack · CIFAR-10adversarial rate: AR roughly 0.26-0.31 at gamma=0.04 vs 0.32 baselineNo clear effect
Adversarial parameter attack on ResNet56 vs VGG19 at same ratioResNet56 · CIFAR-10adversarial rate: ResNet56 gets AR 0.98 at 0.10 but random perturbations are far less effective on ResNet56 than VGG19Mixed
L_0 sparse weight swapping attack instead of L_inf (ablation)VGG19 · CIFAR-10adversarial rate: AR 0.36-0.47 with accuracy drop to 72-76%No clear effect

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