Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Lineage

Runpeng Yu, Xinchao Wang

TL;DR

This paper introduces a novel task known as neural lineage detection, aiming at discovering lineage relationships between parent and child models, and proposes a learning-free and learning-based methods that out-perform the baseline in various learning settings and are adaptable to a variety of visual models.

Abstract

Given a well-behaved neural network, is possible to identify its parent, based on which it was tuned? In this paper, we introduce a novel task known as neural lineage detection, aiming at discovering lineage relationships between parent and child models. Specifically, from a set of parent models, neural lineage detection predicts which parent model a child model has been fine-tuned from. We propose two approaches to address this task. (1) For practical convenience, we introduce a learning-free approach, which integrates an approximation of the finetuning process into the neural network representation similarity metrics, leading to a similarity-based lineage detection scheme. (2) For the pursuit of accuracy, we introduce a learning-based lineage detector comprising encoders and a transformer detector. Through experimentation, we have validated that our proposed learning-free and learning-based methods outperform the baseline in various learning settings and are adaptable to a variety of visual models. Moreover, they also exhibit the ability to trace cross-generational lineage, identifying not only parent models but also their ancestors.

Neural Lineage

TL;DR

This paper introduces a novel task known as neural lineage detection, aiming at discovering lineage relationships between parent and child models, and proposes a learning-free and learning-based methods that out-perform the baseline in various learning settings and are adaptable to a variety of visual models.

Abstract

Given a well-behaved neural network, is possible to identify its parent, based on which it was tuned? In this paper, we introduce a novel task known as neural lineage detection, aiming at discovering lineage relationships between parent and child models. Specifically, from a set of parent models, neural lineage detection predicts which parent model a child model has been fine-tuned from. We propose two approaches to address this task. (1) For practical convenience, we introduce a learning-free approach, which integrates an approximation of the finetuning process into the neural network representation similarity metrics, leading to a similarity-based lineage detection scheme. (2) For the pursuit of accuracy, we introduce a learning-based lineage detector comprising encoders and a transformer detector. Through experimentation, we have validated that our proposed learning-free and learning-based methods outperform the baseline in various learning settings and are adaptable to a variety of visual models. Moreover, they also exhibit the ability to trace cross-generational lineage, identifying not only parent models but also their ancestors.
Paper Structure (47 sections, 2 theorems, 24 equations, 5 figures, 6 tables)

This paper contains 47 sections, 2 theorems, 24 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Deep Model IP Protection
  4. Neural Network Representation Similarity
  5. Neural Network Linearization
  6. Similarity-Based Detection
  7. Notation
  8. Method
  9. Intuitive Explanation
  10. Lineage Detector
  11. Experiments
  12. Classification
  13. Other Tasks and Losses
  14. Conclusion
  15. Further Experiments and Analyses
  16. ...and 32 more sections

Key Result

Proposition 1

Define $\bm{d}_i\triangleq f_p(\bm{x}^{(i)})-f_c(\bm{x}^{(i)})$ to be the difference between the outputs of $f_p$ and $f_c$ for input $\bm{x}^{(i)}$. Let $s(\Bar{f}_p, f_c)$ be the similarity between $\{\Bar{f}_p(\bm{x}^{(i)})\}_{i=1}^N$ and $\{f_c(\bm{x}^{(i)})\}_{i=1}^N$, whose linear approximati where the complete expression of $\zeta_i$ and $\xi_i$ are presented in the Appendix.

Figures (5)

  • Figure 1: Comparison of the execution time and GPU memory Consumption for methods with and without approximation.
  • Figure 2: The influence of approximations and linearized model. \ref{['fig:2:value']} plots the similarity values measured with and without using approximation. \ref{['fig:2:gain']} plots the similarity gains after using the linearized model to get a better anchor.
  • Figure 3: The architecture of the proposed lineage detector. Weights and features are first encoded separately and are then fed into a transformer to obtain the prediction score.
  • Figure 4: Lineage detection result when finetuning ResNet18 on CIFAR10 with different learning rates and iterations.
  • Figure S6: Comparison of the performance when feature position and $\alpha$ vary.

Theorems & Definitions (6)

  • Proposition 1: Similarity Approximation.
  • Remark 1
  • Remark 2
  • Remark 3
  • Remark 4
  • Proposition 2: The Optimality of Measurement