Research map / Architectures and efficiency / Pruning and efficient tuning
Defending Unauthorized Model Merging via Dual-Stage Weight Protection
Wei-Jia Chen, Min-Yan Tsai, Cheng-Yi Lee, Chia-Mu Yu
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 the dual-stage protection before releasing the model | fine-tuned ViT-L-14 · 8 vision datasets (merged, TA) | average accuracy: 84.47 -> 40.26 merged accuracy | Improved |
| apply the protection | fine-tuned ViT-L-14 · 8 vision datasets (standalone) | accuracy: 99.24 -> 98.25 (GTSRB) | No clear effect |
| use the dual-stage method instead of PaRaMS | merged ViT-L-14 · 8 vision datasets (TA) | average accuracy drop: 30.76 -> 52.11 | Improved |
| combine Stage 2 with Stage 1 (ablation) | defense · GTSRB (merged) | accuracy: 64.86 -> 12.91 | Improved |
| combine Stage 1 with Stage 2 (ablation) | defense · MNIST (merged) | accuracy: 13.37 -> 11.35 | Improved |
| combine Stage 1 with Stage 2 (ablation) | defense · SUN397 (merged) | accuracy: 76.38 -> 54.27 | Improved |
| Unmask adaptive attack (subtract protected model) | defense · 8 vision datasets average | merged accuracy: 40.26 -> 64.64 | Hurt |
| gradient-projection adaptive attack | defense · 8 vision datasets average | merged accuracy: 40.26 -> 51.77 | Hurt |
| protection on diffusion model (SD1.5) | defense vs PaRaMS · image generation | visual fidelity: no numbers | Mixed |
Closest papers by meaning
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.90)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.87)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.86)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.86)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.86)
- CABS: Conflict-Aware and Balanced Sparsification for Enhancing Model Merging (ICML 2025 · similarity 0.85)
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.
