Privacy Aware Feature Level Domain Adaptation for Training Deep Vision Models With Private Medical Stethoscope Dataset

Abstract

Source-Free Domain Adaptation (SFDA) is an important research topic in domains with data privacy concerns. Existing SFDA studies have successfully achieved domain adaptation without revealing source domain data, significantly reducing the possibility of privacy leaks. However, complete SFDA (cSFDA), which does not disclose even the model weights of the source domain, has not yet been adequately addressed. Considering the rapidly advancing fields of techniques such as model inversion attacks, we believe that discussions on this cSFDA scenario should be conducted promptly. To perform domain adaptation without revealing both the weights and the data of the source domain, we redefine the domain adaptation process by decomposing it into two stages: information vector extraction and embedding information transfer. In this paper, we propose a novel Spectrogram Secure Domain Adaptation via Encrypted Vector Transfer (SeDA-EVT) pipeline for medical auscultation data, which is achieved by sequentially merging two processes above. We first demonstrated the effectiveness of our information extraction strategy through classification task performance evaluation, showing that our first phase is capable of producing information-rich embeddings. Next, by applying the embedding information transfer to a newly collected clinical lung sound data set from an ER environment, we verified that our proposed pipeline can transfer rich information to the target domain without revealing any source domain components.

Keywords

• Medical auscultation dataset
• self-supervised learning
• source-free domain adaptation