AI Paper Atlas

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

Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training

Yanyun Wang, Li Liu

ICCV 2025 · Adversarial robustness and security › Certified and empirical robustness · approach: Robustness-accuracy trade-off analysis · 3 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 Robust Perception regularizer (RPAT)Consistency-AT · CIFAR-10 (Linf)AutoAttack accuracy: 47.72 -> 48.98Improved
add RPATPGD-AT · CIFAR-10 (Linf)AutoAttack accuracy: 46.74 -> 48.00Improved
add RPATTRADES · CIFAR-10 (Linf)AutoAttack accuracy: 47.62 -> 47.47No clear effect
add RPATMART · CIFAR-10 (Linf)clean accuracy: 77.93 -> 75.44Hurt
add RPATPGD-AT · CIFAR-100 (Linf)AutoAttack accuracy: 25.02 -> 24.88No clear effect
add RPATConsistency-AT · Tiny-ImageNet (Linf)AutoAttack accuracy: 17.60 -> 18.84Improved
swap MSE for KL, JS or cosine loss in Robust Perception term (ablation)PGD-AT+RPAT · CIFAR-10 (Linf)accuracyNo clear effect
use second-to-last layer embedding instead of logits as perception proxy (ablation)PGD-AT+RPAT · CIFAR-10 (Linf)AutoAttack accuracy: 48.00 -> 47.89No clear effect
use alpha 0.5 instead of 0.2, 0.8 or Beta-sampled values (ablation)RPAT · CIFAR-10 (Linf)accuracyImproved
add RPAT to ReBAT (RPAT++)ReBAT · CIFAR-10 (Linf, PreActResNet-18)mean of clean and AA: 66.405 -> 66.815Improved
use stronger training adversary (epsilon 10/255 after first decay) (ablation)ReBAT · CIFAR-10 (Linf, PreActResNet-18)mean of clean and AA: 66.405 -> 64.240Hurt
add RPAT to Consistency-ATConsistency-AT · Tiny-ImageNet (Linf)mean of clean and AA: 32.070 -> 34.290Improved

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