Abstract
Post-deployment adaptation is crucial for maintaining model performance when deployed in new environments or with changing data distributions. However, traditional adaptation approaches often require direct access to source data, which may not be available due to privacy or logistical constraints. In this work, we propose a federated learning approach for post-deployment adaptation that enables effective adaptation while preserving data privacy.
Our method introduces a source-target remote gradient alignment framework that enables adaptation without sharing raw data. The framework consists of three key components: (1) a federated learning protocol that enables collaborative adaptation, (2) a remote gradient alignment mechanism that synchronizes adaptation across domains, and (3) a privacy-preserving adaptation module that maintains data security.
We evaluate our approach on multiple medical imaging datasets with domain shifts, including cross-scanner and cross-institution scenarios. Experimental results demonstrate that our federated adaptation approach significantly improves model performance in new environments compared to traditional methods. The method shows robust performance while maintaining strict privacy standards.
The proposed framework represents a significant advancement in privacy-preserving model adaptation, providing effective post-deployment adaptation that could improve model performance in clinical settings.
BibTeX
@inproceedings{wagner2023post,
title={Post-Deployment Adaptation with Access to Source Data via Federated Learning and Source-Target Remote Gradient Alignment},
author={Wagner, Felix and Li, Zeju and Saha, Pramit and Kamnitsas, Konstantinos},
booktitle={International Workshop on Machine Learning in Medical Imaging (MICCAI-MLMI 2023)},
year={2023},
doi={10.1007/978-3-031-45676-3_26}
}