Research map / Federated learning and privacy / Privacy attacks and secure learning
An Empirical Study on the Resilience of Partial Merging to Model Clone Attacks
Tiantong Wu, Yurong Hao, Wei Yang Bryan Lim
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 |
|---|---|---|---|
| use ModelPirate AS[110] instead of Knockoff | ViT-B/32 clone attack · EuroSAT→DTD | clone accuracy: 7.18% -> 85.15% | Improved |
| use ModelPirate AS[101] instead of JBDA | ViT-B/32 clone attack · DTD→EuroSAT | clone accuracy: 23.48% -> 96.83% | Improved |
| add victim shared part knowledge (AS[100] vs AS[000]) (ablation) | ModelPirate ViT-B/32 · EuroSAT→DTD | clone accuracy: 2.39% -> 37.27% | Improved |
| add known victim training data (AS[001] vs AS[000]) (ablation) | ModelPirate ViT-B/32 · DTD→EuroSAT | clone accuracy: 15.74% -> 60.07% | Improved |
| add known private structure to AS[110] (AS[111]) (ablation) | ModelPirate ViT-B/32 · EuroSAT→DTD | clone accuracy: 85.15% -> 62.89% | Hurt |
| add known private structure to AS[110] (AS[111]) (ablation) | ModelPirate ViT-L/14 · EuroSAT→DTD | clone accuracy: 97.87% -> 98.19% | No clear effect |
| add known private structure to AS[110] (AS[111]) (ablation) | ModelPirate ViT-L/14 · DTD→EuroSAT | clone accuracy: 52.54% -> 63.22% | Improved |
| add victim data to AS[100] (AS[101]) (ablation) | ModelPirate ViT-B/16 · DTD→EuroSAT | clone accuracy: 31.07% -> 95.81% | Improved |
| merge more layers (l >= 7 versus l < 7) (ablation) | ModelPirate ViT-B/32 · EuroSAT→DTD | clone accuracy: clone accuracy surpasses merged accuracy at l = 7 | Hurt |
| increase proportion of known victim data pd from 20% to 100% (ablation) | ModelPirate · EuroSAT→DTD and DTD→EuroSAT | clone accuracy | Improved |
| use deep-shallow structure instead of the original private structure (ablation) | ModelPirate · ViT-B/32 and ViT-L/14 EuroSAT→DTD | clone accuracy: 85.15% -> 62.89% | Improved |
Closest papers by meaning
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.86)
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.86)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.85)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.84)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.84)
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 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.
