AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

From Parameters to Feature Space: Task Arithmetic for Backdoor Mitigation in Model Merging

Zhenqian Zhu, Yamin Hu, Yiya Diao, Weixiang Li, Haodong Li, Wenjian Luo

ICML 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Detecting poisoned samples · 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 LFPM purification instead of plain Task Arithmetic mergeTask Arithmetic merged model · BadMerging (sequence 1)average ASR(T): 98.42 -> 0.49Improved
use LFPM instead of SAM in anti-backdoor optimizationSAM defense · BadMerging (sequence 1)average CA: 40.55 -> 58.67Improved
use LFPM instead of PAMPAM defense · BadMerging (sequence 1)average ASR(N): 21.99 -> 9.85Improved
use LFPM instead of IBVS parameter-space task arithmetic defenseIBVS · BadMerging (sequence 1)average CA: 42.85 -> 58.67Improved
use LFPM instead of SAU adversarial-training defenseSAU · BadMerging (sequence 1)average ASR(T): 63.00 -> 0.49Improved
use LFPM instead of SAM under LoRA attackSAM · LoBAM (sequence 1)average ASR(T): 3.18 -> 3.84No clear effect
use LFPM instead of PAM under LoRA attackPAM · LoBAM (sequence 1)average CA: 52.19 -> 54.45Improved
use LFPM instead of PAM under LoRA attackPAM · LoBAM (sequence 1)average ASR(N): 24.29 -> 26.68Hurt
use LFPM instead of SAU under LoRA attackSAU · LoBAM (sequence 1)average ASR(N): 66.72 -> 26.68Improved
apply adaptive attack (feature coupling plus feature interpolation) against LFPMLFPM · BadMerging (sequence 1)average ASR(T): 0.49 -> 2.38Hurt
apply adaptive attack against SAUSAU · BadMerging (sequence 1)average ASR(T): 63.00 -> 99.20Hurt
apply adaptive attack against SAMSAM · BadMerging (sequence 1)average ASR(T): 10.03 -> 32.30Hurt

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