Secure Domain Adaptation with Multiple Sources

Serban Stan, Mohammad Rostami

Keywords: unsupervised domain adaptation, multi-source domain adaptation, data privacy, source-free adaptation

Abstract: Multi-source unsupervised domain adaptation (MUDA) is a recently explored learning framework within UDA, where the goal is to address the challenge of annotated data scarcity in a target domain via transferring knowledge from multiple source domains with annotated data. When the source domains are distributed, data privacy and security can become a significant concern, e.g., medical domains, yet existing MUDA methods overlook this concern. We develop an algorithm to address MUDA when source domains' data cannot be shared. Our method is based on aligning the distributions of the source and target domains indirectly via internally learned distributions in an intermediate embedding space. Our theoretical analysis supports our approach and extensive empirical results demonstrate our algorithm is effective and compares favorably against existing MUDA methods.

One-sentence Summary: Unsupervised multi source domain adaptation that is source free and keeps privacy between source domains.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2106.12124/code)