AI Paper Atlas

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

Rethinking Robust Adversarial Concept Erasure in Diffusion Models

Qinghong Yin, Yu Tian, Heming Yang, Xiang Chen, Xianlin Zhang, Yue Ming, Xueming Li, Yue Zhang

ECCV 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Diffusion-based adversarial purification · 2 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 semantic-guided prior to single-sample adversarial optimization (ablation)adversarial optimization · adversarial embedding quality (avg of 3 concepts)CLIP-Score: 23.79 -> 24.73Improved
add semantic-guided prior (lambda 0 -> 0.1) (ablation)S-GRACE · NudityUD ASR: 73.94 -> 12.68Improved
raise prior weight lambda from 0.1 to 0.5 (ablation)S-GRACE · NudityUD ASR: 12.68 -> 37.32Hurt
add surrogate-mapping term (alpha 0 -> 1.2) (ablation)S-GRACE · NudityCLIP-Score: 16.04 -> 29.44Improved
add anchor preservation term (beta 0 -> 1.2) (ablation)S-GRACE · NudityCLIP-Score: 16.70 -> 29.44Improved
increase adversarial embeddings P from 8 to 16 (ablation)S-GRACE · NudityUD ASR / utility: UD 12.68 -> 11.97; CLIP 29.44 -> 29.21; FID 15.01 -> 15.96Mixed
use Llama-3 instead of GPT-4 for anchors (ablation)S-GRACE · NudityUD ASR: 12.68 -> 23.94Hurt
use S-GRACE instead of AdvUnlearnS-GRACE vs AdvUnlearn · NudityAvg ASR: 16.38 -> 8.80Improved
use S-GRACE instead of ESDS-GRACE vs ESD · ViolenceAvg ASR: 69.53 -> 22.52Improved
erase Illegal Activity with S-GRACES-GRACE · Illegal ActivityAvg ASR: 76.87 -> 50.68Improved
use S-GRACE against CCE attackS-GRACE · NudityCCE ASR: 57.75 (AdvUnlearn) vs 59.86 (S-GRACE)No clear effect
transfer S-GRACE text encoder to other diffusion modelsS-GRACE · SD v1.5 / DreamShaper / Protogen NudityASR: SD1.5 95.74 -> 22.70; DreamShaper 90.14 -> 43.97; Protogen 83.10 -> 23.24Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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