Research map / Image generation / Text-to-image diffusion
Disrupting Model Merging: A Parameter-Level Defense Without Sacrificing Accuracy
Wei Junhao, Yu Zhe, Jun Sakuma
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 MLP rearrangement and random head scaling (PaRaMS) to the defender model | TA-merged ViT-B-32 · GTSRB/DTD pair | free-rider task accuracy of merged model: 67.82 -> 2.66 | Improved |
| add scaling module to rearrangement-only protection (ablation) | PaRaMS on ViT-L-14 · Cars defender, avg MMP+ | accuracy (lower is better defense): 5.90 -> 4.99 | Improved |
| add rearrangement module to scaling-only protection (ablation) | PaRaMS on ViT-L-14 · Cars defender, avg MMP+ | accuracy (lower is better defense): 10.92 -> 4.99 | Improved |
| add 10% dropout pruning to PaRaMS against an adaptive free-rider (ablation) | PaRaMS ViT-B-32 · SVHN defender merged accuracy | accuracy (lower is better defense): 80.32% -> 64.35% | Improved |
| add 10% dropout pruning to the unmerged defender model (ablation) | PaRaMS ViT-B-32 · SVHN | unmerged accuracy: 97.46% to 96.33% | Hurt |
| free-rider uses adaptive recovery against PaRaMS | PaRaMS ViT-B-32 · SVHN defender | merged accuracy: 7.12% -> 80.32% | Hurt |
| free-rider uses knowledge distillation against PaRaMS+dropout | PaRaMS+dropout · SVHN defender | merged accuracy: 7.01% -> 82.27% | Hurt |
| increase TA merging coefficient from 0.3 to 0.8 (ablation) | PaRaMS ViT-B-32 · EuroSAT defender | MMP+ accuracy: MMP+ at most less than 25% | No clear effect |
| increase number of merged models from 2 to 7 (ablation) | PaRaMS ViT-B-32/ViT-L-14 · MNIST defender | MMP+ accuracy: around 6-7% | No clear effect |
| apply PaRaMS to Stable Diffusion 1.5 merged by TA | PaRaMS · Prompt1 alignment | CLIP alignment: 0.3416 -> 0.1277 | Improved |
| apply PaRaMS to Llama2 full-finetune merged by TA | PaRaMS · Emotion | accuracy of merged model: 97.6 -> 21.4 | Improved |
| apply scaling-only PaRaMS to PEM-TA merged LoRA | PaRaMS · Emotion | accuracy of merged model: 96.1 -> 52.7 | Improved |
Closest papers by meaning
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.88)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.86)
- Scalable Model Merging with Progressive Layer-wise Distillation (ICML 2025 · similarity 0.85)
- CABS: Conflict-Aware and Balanced Sparsification for Enhancing Model Merging (ICML 2025 · similarity 0.85)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.85)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.84)
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.
