Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy-Preserving Federated Unsupervised Domain Adaptation for Regression on Small-Scale and High-Dimensional Biological Data

Cem Ata Baykara, Ali Burak Ünal, Nico Pfeifer, Mete Akgün

TL;DR

Freda tackles domain shifts in small, high-dimensional biological data under strict privacy constraints by enabling privacy-preserving federated unsupervised domain adaptation for regression. It uniquely federates Gaussian Process Regressors through randomized encoding and secure aggregation (including zero-sum masking and the FLAKE framework) to compute necessary kernel quantities ($K$, $K_*$) and predictions without revealing raw data, followed by a target-driven feature weighting and federated weighted elastic net training with optimal lambda predicted via a prior-knowledge approach. The method is evaluated on age prediction from DNA methylation data and achieves performance comparable to centralized state-of-the-art baselines while preserving complete data privacy, demonstrating its potential for cross-institution collaboration in biomedical settings. Overall, freda provides a scalable, data-efficient framework for secure domain adaptation in high-dimensional biology, extending the applicability of GP-based modeling to distributed privacy-sensitive scenarios.

Abstract

Machine learning models often struggle with generalization in small, heterogeneous datasets due to domain shifts caused by variations in data collection and population differences. This challenge is particularly pronounced in biological data, where data is high-dimensional, small-scale, and decentralized across institutions. While federated domain adaptation methods (FDA) aim to address these challenges, most existing approaches rely on deep learning and focus on classification tasks, making them unsuitable for small-scale, high-dimensional applications. In this work, we propose freda, a privacy-preserving federated method for unsupervised domain adaptation in regression tasks. Unlike deep learning-based FDA approaches, freda is the first method to enable the federated training of Gaussian Processes to model complex feature relationships while ensuring complete data privacy through randomized encoding and secure aggregation. This allows for effective domain adaptation without direct access to raw data, making it well-suited for applications involving high-dimensional, heterogeneous datasets. We evaluate freda on the challenging task of age prediction from DNA methylation data, demonstrating that it achieves performance comparable to the centralized state-of-the-art method while preserving complete data privacy.

Privacy-Preserving Federated Unsupervised Domain Adaptation for Regression on Small-Scale and High-Dimensional Biological Data

TL;DR

Freda tackles domain shifts in small, high-dimensional biological data under strict privacy constraints by enabling privacy-preserving federated unsupervised domain adaptation for regression. It uniquely federates Gaussian Process Regressors through randomized encoding and secure aggregation (including zero-sum masking and the FLAKE framework) to compute necessary kernel quantities (, ) and predictions without revealing raw data, followed by a target-driven feature weighting and federated weighted elastic net training with optimal lambda predicted via a prior-knowledge approach. The method is evaluated on age prediction from DNA methylation data and achieves performance comparable to centralized state-of-the-art baselines while preserving complete data privacy, demonstrating its potential for cross-institution collaboration in biomedical settings. Overall, freda provides a scalable, data-efficient framework for secure domain adaptation in high-dimensional biology, extending the applicability of GP-based modeling to distributed privacy-sensitive scenarios.

Abstract

Machine learning models often struggle with generalization in small, heterogeneous datasets due to domain shifts caused by variations in data collection and population differences. This challenge is particularly pronounced in biological data, where data is high-dimensional, small-scale, and decentralized across institutions. While federated domain adaptation methods (FDA) aim to address these challenges, most existing approaches rely on deep learning and focus on classification tasks, making them unsuitable for small-scale, high-dimensional applications. In this work, we propose freda, a privacy-preserving federated method for unsupervised domain adaptation in regression tasks. Unlike deep learning-based FDA approaches, freda is the first method to enable the federated training of Gaussian Processes to model complex feature relationships while ensuring complete data privacy through randomized encoding and secure aggregation. This allows for effective domain adaptation without direct access to raw data, making it well-suited for applications involving high-dimensional, heterogeneous datasets. We evaluate freda on the challenging task of age prediction from DNA methylation data, demonstrating that it achieves performance comparable to the centralized state-of-the-art method while preserving complete data privacy.

Paper Structure

This paper contains 38 sections, 14 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Overview
  4. Dataset and Pre-Processing
  5. Baselines
  6. Wenda Baseline
  7. Non-Adaptive Baseline
  8. Setup for Freda
  9. Data Distribution
  10. Setup for Weighted Elastic Net Models
  11. Parameter Selection in Freda
  12. Experiments
  13. Discussion
  14. Related Work
  15. Domain Adaptation
  16. ...and 23 more sections

Figures (3)

  • Figure 1: Overview of the freda framework, in which multiple source domain clients collaborate to perform domain adaptation on the target domain data with the assistance of an aggregator.
  • Figure 2: Mean absolute error per target tissue, as well as on full target data for the non-adaptive and non-federated baseline en-ls, wenda-pn with $k=3$, and freda with $k=3$, across 2, 4, and 8 source parties.
  • Figure 3: Predicted versus true chronological age under various settings. Figures (a), (b), and (c) correspond to freda with $k=3$ for 2, 4, and 8 source parties, respectively. Predictions are averaged over all splits where the tissue of interest was included in the evaluation set, as well as over 5 different distributions for each setting. Panels (d) and (e) correspond to en-ls and wenda-pn, respectively. For en-ls, predictions are averaged over 10 runs of 10-fold cross-validation, while for wenda-pn, predictions are averaged over all splits where the tissue of interest was included in the evaluation set.