Research map / Adversarial robustness and security / Backdoor and poisoning
On the Vulnerability of Parameter-Level Defenses to Model Merging
Kuangpu Guo, Jian Liang, Qingyan Zheng, Yu Yongcan, Zilei Wang, Ran He, Tieniu Tan
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 |
|---|---|---|---|
| apply AGA attack to MergeLock-protected models | Task Arithmetic merge · ViT-B/32 protected task | avg accuracy: 5.10 -> 66.39 | Improved |
| apply AGA attack to MergeLock-protected models | LOT Merging · ViT-B/32 protected task | avg accuracy: 4.85 -> 78.36 | Improved |
| apply AGA attack to MergeBarrier-protected models | Task Arithmetic merge · ViT-B/32 protected task | avg accuracy: 29.32 -> 60.80 | Improved |
| apply AGA attack to Params-D-protected models | Qwen2-7B · Alpaca/GSM8K/MBPP | avg score: 3.07 -> 47.79 | Improved |
| apply AGA attack to MergeLock-protected models | ViT-L/14 · protected task | avg accuracy: 4.27 -> 82.72 | Improved |
| combine Params, MergeLock and MergeBarrier defenses then attack with AGA | combined defense · ViT-B/32 protected task | avg accuracy: 4.16 -> 60.66 | Improved |
| apply AGA to MergeLock-protected merged model | Task Arithmetic merge · cross-task average utility | avg accuracy: 4.92 -> 67.17 | Improved |
| replace MergeGuard with ARF protection | ARF · ViT-B/32 protected task under AGA | avg accuracy: 50.56 -> 27.42 | Improved |
| replace MergeGuard with ARF protection | ARF · ViT-L/14 protected task under AGA | avg accuracy: 63.60 -> 29.26 | Improved |
| replace MergeGuard with ARF protection | ARF · ViT-B/32 standalone accuracy | avg accuracy: 86.31 -> 88.11 | Improved |
| increase ARF expansion ratio rho from 0.02 to 0.08 (ablation) | ARF · ViT-B/32 protected under AGA | avg accuracy: 28.75 -> 26.36 | Mixed |
| increase ARF lambda_dist from 0.7 to 1.3 (ablation) | ARF · ViT-B/32 protected under AGA | avg accuracy: 29.26 -> 26.17 | Mixed |
Closest papers by meaning
- Disrupting Model Merging: A Parameter-Level Defense Without Sacrificing Accuracy (ICCV 2025 · similarity 0.90)
- Defending Unauthorized Model Merging via Dual-Stage Weight Protection (CVPR 2026 · similarity 0.88)
- AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion (NeurIPS 2026 · similarity 0.88)
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.86)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.83)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.83)
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.
