Trustworthy Transfer Learning: A Survey

TL;DR

This survey tackles trustworthy transfer learning by integrating two core dimensions: knowledge transferability and knowledge trustworthiness across IID and non-IID settings, varied data modalities, and real-world applications. It surveys theoretical foundations (discrepancy metrics, margin-based and f-divergence frameworks), transferability estimation, and modality-specific transfer analyses (graphs, text, time series), followed by privacy, robustness, fairness, transparency, and sustainability considerations in transfer scenarios. The paper also covers practical aspects such as federation, test-time and source-free adaptations, and foundation-model concerns, including negative transfer and cross-modal and physics-informed transfer. By outlining open questions and future directions, the authors advocate for a unified, responsible framework that improves target performance while maintaining rigorous trustworthiness in diverse deployment contexts.

Abstract

Transfer learning aims to transfer knowledge or information from a source domain to a relevant target domain. In this paper, we understand transfer learning from the perspectives of knowledge transferability and trustworthiness. This involves two research questions: How is knowledge transferability quantitatively measured and enhanced across domains? Can we trust the transferred knowledge in the transfer learning process? To answer these questions, this paper provides a comprehensive review of trustworthy transfer learning from various aspects, including problem definitions, theoretical analysis, empirical algorithms, and real-world applications. Specifically, we summarize recent theories and algorithms for understanding knowledge transferability under (within-domain) IID and non-IID assumptions. In addition to knowledge transferability, we review the impact of trustworthiness on transfer learning, e.g., whether the transferred knowledge is adversarially robust or algorithmically fair, how to transfer the knowledge under privacy-preserving constraints, etc. Beyond discussing the current advancements, we highlight the open questions and future directions for understanding transfer learning in a reliable and trustworthy manner.

Figures (11)

• Figure 1: A motivating example of trustworthy transfer learning
• Figure 2: Overview of trustworthy transfer learning (best viewed in color)
• Figure 3: Evaluation of transferability between the pre-trained source model and the target data: (a) Transferability scores select the best source model for the target data given a large pool of pre-trained source models. (b) Transferability scores identify the most suitable application domains/tasks for a source model.
• Figure 4: Illustration of distribution shifts in characterizing graph transferability. The node distribution within the graph can be represented by $\mathrm{P}(X, G, Y)$ where $X, G, Y$ denote the input node attributes, topology structure, and output class labels, respectively. The color of nodes indicates the class labels $Y$ (blue or green).
• Figure 6: Illustration of time series analysis under distribution shifts: (a) time series forecasting, and (b) time series classification

Theorems & Definitions (13)

• definition 1: Trustworthy Transfer Learning
• definition 2: Integral Probability Metric muller1997integral
• theorem 1: Unified Generalization Bound
• definition 3: Task Diversity task_diversity
• definition 4: Transferability Measure NCE
• definition 5: Time Series Transferability for Forecasting DAIN
• definition 6: Time Series Transferability for Classification VRADA
• definition 7: Hypothesis Transfer KuzborskijO13
• definition 8: Federated Transfer
• definition 9: Fairness Transfer schumann2019transferchen2022fairness