AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

Zero-Sacrifice Persistent-Robustness Adversarial Defense for Pre-Trained Encoders

Zhuxin Lei, Ziyuan Yang, Yi Zhang

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Defenses against universal perturbations · 0 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 adversarial fine-tuning to the dual-branch encoder (ablation)ZePAD · ANIMALS10 (CIFAR10-pretrained DINO)robust accuracy: 59.40 -> 74.71Improved
add the MPAE branch (ablation)ZePAD · ANIMALS10 (CIFAR10-pretrained DINO)robust accuracy: 70.35 -> 74.71Improved
add the BMP branch (ablation)ZePAD · ANIMALS10 (CIFAR10-pretrained DINO)benign accuracy: 84.51 -> 91.66Improved
add the BMP branch (ablation)ZePAD · ANIMALS10 (CIFAR10-pretrained DINO)attack success rate: 19.36 -> 21.99Hurt
add the BMP branch (ablation)ZePAD · ANIMALS10 (CIFAR10-pretrained DINO)robust accuracy: 72.47 -> 74.71Improved
apply ZePAD to the SSL encoder instead of plain baselineDINO encoder · ANIMALS10 (CIFAR10-pretrained DINO)robust accuracy: 35.46 -> 74.71Improved
use ZePAD instead of baseline, CIFAR10-pretrained, SVHNSSL encoders avg · SVHN (CIFAR10-pretrained)robust accuracy: 13.11 -> 86.97Improved
use ZePAD instead of baseline, CIFAR10-pretrained, ANIMALS10SSL encoders avg · ANIMALS10 (CIFAR10-pretrained)benign accuracy: 88.52 -> 98.14Improved
use ZePAD instead of baseline, CIFAR10-pretrained, STL10SSL encoders avg · STL10 (CIFAR10-pretrained)benign accuracy: 82.63 -> 82.51No clear effect
use ZePAD instead of baseline, CIFAR10-pretrained, CIFAR10SSL encoders avg · CIFAR10 (CIFAR10-pretrained)benign accuracy: 93.34 -> 93.57No clear effect
use ZePAD vs Gen-AF defenseSimCLR · STL10 (ImageNet-pretrained)benign accuracy: 68.69 -> 82.07Improved
use ZePAD vs Gen-AF defenseSimCLR · STL10 (ImageNet-pretrained)robust accuracy under AdvEncoder: 62.28 -> 73.26Improved

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