Research map / Detection and segmentation / Open-set and few-shot detection
Neural Lineage
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.
Introduces neural lineage detection, the task of identifying which parent model a fine-tuned child model came from, with a learning-free linearization-based similarity method and a learning-based transformer lineage detector.
Problem
Given a child model, identify which of a set of candidate parent models it was fully fine-tuned from, without relying on watermarks or fingerprints.
Approach
Approximates the fine-tuning process via first-order Taylor linearization (NTK-style) and folds it into model similarity metrics via a single-backprop approximation; a learning-based detector encodes parent/child weights and features and classifies with a transformer.
Key results
- Learning-based lineage detector outperforms the others by an average of 36.26% accuracy on classification tasks.
- Learning-free linearized similarity improves over the baseline by an average of 3.17%, with a maximum gain of 28%.
- Lineage detector reaches about 99% accuracy on several FC and ResNet18 settings (e.g., 99.11% CIFAR10 on ResNet18 setting in Table 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | ResNet18 lineage detection · CIFAR10 | accuracy: 94.64±0.63 -> 99.87±0.07 | Helped |
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | ResNet18 lineage detection · CIFAR10 imbalanced | accuracy: 95.84±0.99 -> 97.01±0.41 | Helped |
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | ResNet18 lineage detection · CIFAR10 5-shot | accuracy: 99.89±0.62 -> 99.91±0.52 | No clear effect |
| Add the fine-tuning approximation to l1 similarity (ablation) | ResNet18 lineage detection · FMNIST | accuracy: 90.37±0.82 -> 98.81±0.41 | Helped |
| Add the fine-tuning approximation to l1 similarity (ablation) | ResNet18 lineage detection · CIFAR10 | accuracy: 90.19±1.33 -> 94.64±0.63 | Helped |
| Add the fine-tuning approximation to l1 similarity (ablation) | ResNet18 lineage detection · CIFAR10 imbalanced | accuracy: 87.50±1.11 -> 95.84±0.99 | Helped |
| Add the fine-tuning approximation to CKA (ablation) | ResNet18 lineage detection · CIFAR10 | accuracy: 45.83±0.81 -> 47.62±0.94 | No clear effect |
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | lineage detection · Object detection | accuracy: 96.42±1.23 -> 99.28±0.72 | Helped |
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | lineage detection · Segmentation | accuracy: 35.41±2.29 -> 37.56±4.04 | Helped |
| Use the learning-based Lineage Detector instead of l2 with approximation | lineage detection · Continual learning | accuracy: 99.83±0.21 -> 99.18±0.63 | Hurt |
| Use the learning-based Lineage Detector instead of the strongest similarity-based baseline | lineage detection · Knowledge distillation | accuracy: 94.03±1.23 -> 99.07±0.07 | Helped |
| Increase fine-tuning learning rate or iteration count (ablation) | ResNet18 lineage detection on CIFAR10 · CIFAR10 | accuracy | Hurt |
Surprising findings
- Norm distances (l1, l2) generally outperform advanced neural network similarity metrics such as CKA and distance correlation for lineage detection.
- Lineage detection accuracy declines as fine-tuning learning rate or iterations increase, though the learning-based method is more robust.
Limitations
- Assumes full-parameter fine-tuning only.
- Learning-based detector requires training on labeled parent-child pairs.
- Accuracy drops as learning rate, iterations, or generation gap increase.
Experimental setup
Datasets: MNIST CIFAR100 ImageNet FMNIST EMNIST-Letters EMNIST-Balanced CIFAR10 SVHN Pet DTD PASCAL
Benchmarks: CIFAR10 FMNIST EMNIST SVHN DTD Pet PASCAL detection/segmentation
Metrics: lineage detection accuracy
Baselines: CKA DC l1 distance l2 distance l-infinity distance lp distance log-sum-exp similarity
Builds on: Neural Tangent Kernel Centered Kernel Alignment Distance Correlation EWC
Closest papers by meaning
- Model Lineage Closeness Analysis (AAAI 2025 · similarity 0.85)
- What Do You See in Common? Learning Hierarchical Prototypes over Tree-of-Life to Discover Evolutionary Traits (ICLR 2025 · similarity 0.83)
- Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning (ICML 2025 · similarity 0.82)
- Unsupervised Model Tree Heritage Recovery (ICLR 2025 · similarity 0.80)
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes (NeurIPS 2026 · similarity 0.79)
- PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders (ICLR 2025 · similarity 0.79)
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.