Research map / Learning theory and algorithms / Prediction and decision losses
Model Lineage Closeness Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a modification-type and task-agnostic method that measures model lineage closeness from decision-boundary similarity (mean adversarial distance and matching rate over efficiently sampled test points), and builds a 136-model benchmark covering modification types.
Problem
Existing model lineage determination methods are not robust to high-impact modifications, require matching output dimensions, and cannot quantify how closely two models are related (lineage closeness, i.e., degree of modification).
Approach
Lineage closeness is measured as the similarity of decision boundaries: data points sampled from the training set are probed with gradient-direction adversarial distances to both models, combined with a matching rate of outputs, s = (1 - mean distance difference) * matching rate. Transfer-learned models with different tasks are handled by first transferring the source model on the compared model's
Key results
- The proposed method achieves 97% lineage determination accuracy (193/198 pairs), versus 79% to 88% for baselines such as MeTaFinger (88%) and CAE (86%).
- Only five misclassified pairs, all distilled models, which the authors attribute to substantially different decision boundaries.
- Lineage closeness scores separate lineage from no-lineage models, and models modified more heavily get lower scores; the test-set generation step adds about 800 s and lineage measurement takes about 10 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed lineage determination method instead of DEEPJUDGE | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 165(83%) -> 193(97%) | Helped |
| use the proposed method instead of MeTaFinger | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 174(88%) -> 193(97%) | Helped |
| use the proposed method instead of CAE | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 170(86%) -> 193(97%) | Helped |
| use the proposed method instead of IPGuard | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 157(79%) -> 193(97%) | Helped |
| use the proposed method instead of ModelDiff | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 147(74%) -> 193(97%) | Helped |
| use the proposed method instead of SAC | Model lineage determination · Model lineage benchmark (198 model pairs) | determination accuracy: 125(63%) -> 193(97%) | Helped |
| apply greater-magnitude model modifications | Model lineage closeness measurement · Lineage models and no-lineage models | lineage closeness score | Hurt |
| use indirect lineage rather than direct lineage | Model lineage closeness measurement · Indirect lineage | lineage closeness score | Hurt |
| increase pruning from 10% to 30% of weights | Model lineage closeness measurement · Pruned models | lineage closeness score | Hurt |
| distill the source model | Source model · Distilled models | decision-boundary similarity | Hurt |
Surprising findings
- Only five misclassified pairs of the proposed method were all distilled models, which the authors attribute to large decision-boundary disparities from their source models.
Limitations
- Distilled models are misclassified because their decision boundaries differ substantially from the source.
- Test set generation is costly (about 800 s) and relies on generating many fine-tuned and no-lineage models.
- Evaluation is limited to image classification with small CNN architectures and four datasets.
Experimental setup
Datasets: MNIST CIFAR-10 Oxford Flowers 102 Stanford Dogs 120 ImageNet
Benchmarks: Model lineage closeness benchmark (136 models, 198 model pairs)
Metrics: lineage determination accuracy time cost (sec) lineage closeness score
Baselines: IPGuard CAE ModelDiff SAC MeTaFinger DEEPJUDGE
Builds on: ModelDiff DEEPJUDGE IPGuard CAE MeTaFinger
Closest papers by meaning
- Neural Lineage (CVPR 2024 · similarity 0.85)
- Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning (ICML 2025 · similarity 0.82)
- Prescriptive Scaling Reveals the Evolution of Language Model Capabilities (ICML 2026 · similarity 0.80)
- We Should Chart an Atlas of All the World's Models (NeurIPS 2025 · similarity 0.80)
- What Do You See in Common? Learning Hierarchical Prototypes over Tree-of-Life to Discover Evolutionary Traits (ICLR 2025 · similarity 0.80)
- Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One (NeurIPS 2025 · similarity 0.80)
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.