Source Data-Absent Unsupervised Domain Adaptation Through Hypothesis Transfer and Labeling Transfer.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) typically requires access to source data, limiting its application in privacy-sensitive scenarios.
• Existing UDA methods are not suitable when source data is confidential and cannot be shared.

Purpose of the Study:

• To develop a novel UDA approach that adapts models using only a pre-trained source classification model, without direct access to source data.
• To address the challenge of knowledge transfer in realistic settings where data privacy is a concern.

Main Methods:

• Proposes Source Hypothesis Transfer (SHOT), a method that learns target domain feature extractors by aligning target features with a frozen source classification model.
• Employs information maximization and self-supervised learning within SHOT to align target features with source data representations.
• Introduces SHOT++, an extension incorporating a labeling transfer strategy using semi-supervised learning on confident target predictions.

Main Results:

• SHOT and SHOT++ demonstrate superior or comparable performance to state-of-the-art methods on digit classification and object recognition tasks.
• The proposed methods effectively adapt models in visual domain adaptation problems under privacy constraints.
• Experimental results validate the efficacy of utilizing a source model hypothesis for knowledge transfer.

Conclusions:

• SHOT and SHOT++ offer effective solutions for unsupervised domain adaptation in privacy-preserving scenarios.
• The approaches successfully transfer knowledge from source models to unlabeled target domains without requiring source data.
• These methods advance the field of UDA by enabling adaptation in previously inaccessible confidential data settings.