AI Paper Atlas

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

AEGIS: Adversarial Target–Guided Retention-Data-Free Robust Concept Erasure from Diffusion Models

Fengpeng Li, Kemou Li, Qizhou Wang, Bo Han, Jiantao Zhou

ICLR 2026 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency · 4 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 erasure target (AET) (ablation)AEGIS · Nudity, UnlearnDiffAtkASR: 52.11% -> 8.45%Improved
add parameter regularization instead of COCO retention data (ablation)AEGIS · NudityFID: 18.15 -> 17.43Improved
add directional gradient rectification (ablation)AEGIS · NudityASR: 26.24% -> 8.45%Improved
add directional gradient rectification (ablation)AEGIS · NudityFID: 19.84 -> 17.43Improved
use dynamic omega rather than fixed omega=1 (ablation)AEGIS · NudityASR: 14.08% -> 8.45%Improved
apply AET to ESD (ablation)ESD · Nudity, UnlearnDiffAtkASR: 76.05% -> 26.06%Improved
apply AET to AdvUnlearn (ablation)AdvUnlearn · Nudity, UnlearnDiffAtkASR: 64.79% -> 5.63%Improved
add GRP to AdvUnlearn (AEGIS w/o AET) (ablation)AdvUnlearn · Nudity, UnlearnDiffAtkFID: 19.34 -> 17.54Improved
move base model from SD v1.4 to SD v2.1AEGIS · Van GoghASR1: 36% -> 40%Hurt
AEGIS vs SH on ChurchAEGIS · ChurchFID: 68.02 -> 19.06Improved

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