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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add LFPM purification instead of plain Task Arithmetic merge | Task Arithmetic merged model · BadMerging (sequence 1) | average ASR(T): 98.42 -> 0.49 | Improved |
| use LFPM instead of SAM in anti-backdoor optimization | SAM defense · BadMerging (sequence 1) | average CA: 40.55 -> 58.67 | Improved |
| use LFPM instead of PAM | PAM defense · BadMerging (sequence 1) | average ASR(N): 21.99 -> 9.85 | Improved |
| use LFPM instead of IBVS parameter-space task arithmetic defense | IBVS · BadMerging (sequence 1) | average CA: 42.85 -> 58.67 | Improved |
| use LFPM instead of SAU adversarial-training defense | SAU · BadMerging (sequence 1) | average ASR(T): 63.00 -> 0.49 | Improved |
| use LFPM instead of SAM under LoRA attack | SAM · LoBAM (sequence 1) | average ASR(T): 3.18 -> 3.84 | No clear effect |
| use LFPM instead of PAM under LoRA attack | PAM · LoBAM (sequence 1) | average CA: 52.19 -> 54.45 | Improved |
| use LFPM instead of PAM under LoRA attack | PAM · LoBAM (sequence 1) | average ASR(N): 24.29 -> 26.68 | Hurt |
| use LFPM instead of SAU under LoRA attack | SAU · LoBAM (sequence 1) | average ASR(N): 66.72 -> 26.68 | Improved |
| apply adaptive attack (feature coupling plus feature interpolation) against LFPM | LFPM · BadMerging (sequence 1) | average ASR(T): 0.49 -> 2.38 | Hurt |
| apply adaptive attack against SAU | SAU · BadMerging (sequence 1) | average ASR(T): 63.00 -> 99.20 | Hurt |
| apply adaptive attack against SAM | SAM · BadMerging (sequence 1) | average ASR(T): 10.03 -> 32.30 | Hurt |
Closest papers by meaning
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.94)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.86)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.86)
- CABS: Conflict-Aware and Balanced Sparsification for Enhancing Model Merging (ICML 2025 · similarity 0.86)
- Localizing Task Information for Improved Model Merging and Compression (ICML 2024 · similarity 0.86)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.86)
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.
