Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Robust Non-Transferable Learning: A Survey and Benchmark

Ziming Hong, Yongli Xiang, Tongliang Liu

TL;DR

This work provides the first systematic survey of non-transferable learning (NTL) and introduces NTLBench, a unified benchmark to evaluate NTL methods across datasets, architectures, and post-training attacks. It highlights robustness gaps in existing NTL approaches, showing that current methods struggle to maintain non-transferability under source-domain and target-domain fine-tuning and SFDA. The paper proposes a formal problem setup, two NTL settings (target-specified and source-only), and classifies techniques into output-space and feature-space regularizations, plus data-generation and augmentation strategies. By assessing 5 SOTA NTL methods on 9 datasets with 15 attacks, it demonstrates significant limitations and emphasizes the need for more robust NTL designs for safe and trustworthy deployment. The work also outlines practical applications (ownership verification, applicability authorization, safety alignment) and future directions to advance robustness and cross-modal generalization of NTL.

Abstract

Over the past decades, researchers have primarily focused on improving the generalization abilities of models, with limited attention given to regulating such generalization. However, the ability of models to generalize to unintended data (e.g., harmful or unauthorized data) can be exploited by malicious adversaries in unforeseen ways, potentially resulting in violations of model ethics. Non-transferable learning (NTL), a task aimed at reshaping the generalization abilities of deep learning models, was proposed to address these challenges. While numerous methods have been proposed in this field, a comprehensive review of existing progress and a thorough analysis of current limitations remain lacking. In this paper, we bridge this gap by presenting the first comprehensive survey on NTL and introducing NTLBench, the first benchmark to evaluate NTL performance and robustness within a unified framework. Specifically, we first introduce the task settings, general framework, and criteria of NTL, followed by a summary of NTL approaches. Furthermore, we emphasize the often-overlooked issue of robustness against various attacks that can destroy the non-transferable mechanism established by NTL. Experiments conducted via NTLBench verify the limitations of existing NTL methods in robustness. Finally, we discuss the practical applications of NTL, along with its future directions and associated challenges.

Toward Robust Non-Transferable Learning: A Survey and Benchmark

TL;DR

This work provides the first systematic survey of non-transferable learning (NTL) and introduces NTLBench, a unified benchmark to evaluate NTL methods across datasets, architectures, and post-training attacks. It highlights robustness gaps in existing NTL approaches, showing that current methods struggle to maintain non-transferability under source-domain and target-domain fine-tuning and SFDA. The paper proposes a formal problem setup, two NTL settings (target-specified and source-only), and classifies techniques into output-space and feature-space regularizations, plus data-generation and augmentation strategies. By assessing 5 SOTA NTL methods on 9 datasets with 15 attacks, it demonstrates significant limitations and emphasizes the need for more robust NTL designs for safe and trustworthy deployment. The work also outlines practical applications (ownership verification, applicability authorization, safety alignment) and future directions to advance robustness and cross-modal generalization of NTL.

Abstract

Over the past decades, researchers have primarily focused on improving the generalization abilities of models, with limited attention given to regulating such generalization. However, the ability of models to generalize to unintended data (e.g., harmful or unauthorized data) can be exploited by malicious adversaries in unforeseen ways, potentially resulting in violations of model ethics. Non-transferable learning (NTL), a task aimed at reshaping the generalization abilities of deep learning models, was proposed to address these challenges. While numerous methods have been proposed in this field, a comprehensive review of existing progress and a thorough analysis of current limitations remain lacking. In this paper, we bridge this gap by presenting the first comprehensive survey on NTL and introducing NTLBench, the first benchmark to evaluate NTL performance and robustness within a unified framework. Specifically, we first introduce the task settings, general framework, and criteria of NTL, followed by a summary of NTL approaches. Furthermore, we emphasize the often-overlooked issue of robustness against various attacks that can destroy the non-transferable mechanism established by NTL. Experiments conducted via NTLBench verify the limitations of existing NTL methods in robustness. Finally, we discuss the practical applications of NTL, along with its future directions and associated challenges.

Paper Structure

This paper contains 44 sections, 1 equation, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Problem Setup
  4. General Framework of NTL
  5. NTL Criteria
  6. Non-transferability performance.
  7. Post-training robustness.
  8. Approaches for NTL
  9. Target-Specified NTL
  10. Output space regularization.
  11. Feature space regularization.
  12. Source-Only NTL
  13. Adversarial domain generation.
  14. Strong image augmentation.
  15. Perturbation-based method.
  16. ...and 29 more sections

Figures (2)

  • Figure 1: We systematically review non-transferable learning (NTL) and introduce NTLBench, an unified framework for benchmarking NTL. This figure compares 5 methods (NTL, CUTI-domain, H-NTL, SOPHON, CUPI-domain) on CIFAR & STL with VGG-13, evaluating pre-training performance and robustness against 5 source domain fine-tuning attacks, 4 target domain fine-tuning attacks, and 6 source-free domain adaptation attacks (higher value means better robustness). NTLBench will be released soon at https://github.com/tmllab/NTLBench.
  • Figure 2: Comparison of (a) supervised learning (SL), (b) target-specified non-transferable learning (NTL), and (c) source-only NTL.