Source-linked AI summary
Differential Privacy-enabled Federated Learning for Sensitive Health Data
Olivia Choudhury, Aris Gkoulalas-Divanis, Theodoros Salonidis, Issa Sylla, Yoonyoung Park, Grace Hsu, Amar Das
TL;DR
Healthcare machine learning faces distributed data silos and privacy constraints that complicate centralized aggregation of sensitive patient data. This paper combines federated learning with distributed differential privacy and evaluates it on two real-world applications involving 1 million patients. Federated learning approached centralized utility, but differential privacy substantially reduced model performance because of added noise.
Problem
Healthcare data are distributed across silos and restricted by privacy requirements, limiting centralized learning from sensitive patient records.
Method
The paper extends federated learning with distributed differential privacy, training global models from local healthcare data without sharing raw records.
Results
Federated learning achieved comparable utility to centralized learning for adverse-drug-reaction and mortality prediction, while differential privacy significantly deteriorated federated-model utility.
Takeaways & Limitations
The evaluation demonstrates federated learning’s feasibility for real-world healthcare applications but motivates alternative approaches to privacy-preserving federated learning.
Takeaways & Limitations
Performance-preserving differential privacy may require about 1000 sites, an unrealistic scale for typical healthcare applications.
Abstract
from arXiv · showhide
Leveraging real-world health data for machine learning tasks requires addressing many practical challenges, such as distributed data silos, privacy concerns with creating a centralized database from person-specific sensitive data, resource constraints for transferring and integrating data from multiple sites, and risk of a single point of failure. In this paper, we introduce a federated learning framework that can learn a global model from distributed health data held locally at different sites. The framework offers two levels of privacy protection. First, it does not move or share raw data across sites or with a centralized server during the model training process. Second, it uses a differential privacy mechanism to further protect the model from potential privacy attacks. We perform a comprehensive evaluation of our approach on two healthcare applications, using real-world electronic health data of 1 million patients. We demonstrate the feasibility and effectiveness of the federated learning framework in offering an elevated level of privacy and maintaining utility of the global model.
1 Introduction
The paper addresses healthcare data silos and privacy constraints by extending federated learning with differential privacy. It evaluates this framework on real-world health data and finds a privacy–utility trade-off.
- Healthcare data are distributed across silos and restricted by HIPAA and GDPR, while centralized learning requires aggregating sensitive data.
- Federated learning trains a global model from local data without sharing raw data across sites.Sites train shared models locally and send parameter updates to an aggregation server.
- The framework adds distributed differential privacy because federated learning can remain vulnerable to inference attacks.Prior differential-privacy performance had not been investigated for healthcare applications.
- Two real-world health datasets comprising electronic health records and administrative claims data covered 1 million patients.
- Federated learning without differential privacy achieved performance close to centralized learning, whereas differential privacy reduced predictive capability through added noise.
2 Methods
The methods define local healthcare datasets and iterative federated optimization, then introduce differential privacy as a formal privacy guarantee. The framework explores noise-based privacy mechanisms, including objective perturbation, for distributed learning.
- Federated learning: The classification setup represents features xk in a feature space X and labels yk in Y := {−1, 1}.
- Federated learning: Positive and negative examples are partitioned into X+ and X− according to their labels.X+ contains examples with yk = +1, while X− contains examples with yk = −1.
- Federated learning: For N sites, each site trains a shared global model on its local dataset Di and computes an average gradient during local optimization.
- Federated learning: The server computes a weighted average of local parameter updates and repeats the process until a convergence criterion is satisfied.
- Differential Privacy: Differential privacy requires similar output distributions for datasets differing by one record, limiting the influence of any individual record.
- Differential Privacy: The paper considers input, output, exponential-mechanism, and objective perturbation, emphasizing objective perturbation because prior work provides theoretical performance guarantees.
3 Evaluation
The evaluation uses two real-world healthcare prediction tasks and compares centralized learning, federated learning, and federated learning with differential privacy. Across both datasets and all three classifiers, increasing ϵ improves utility, while differential privacy substantially reduces federated-model utility.
- 3.1 Use cases and data preparation: The evaluation covers adverse drug reaction prediction with LCED and in-hospital mortality prediction with MIMIC III.LCED contains 921,167 samples, while the MIMIC III cohort contains 21,139 ICU stays.
- 3.2 Experimental setup: The models use perceptron, support vector machine, and logistic regression, with utility measured by F1 score in a 10-site federated setup.Training used 70% of the data with 5-fold cross-validation.
- 3.3 Comparative analysis: As ϵ increases from 0.01 to 0.5, privacy decreases while F1-score utility improves consistently across both datasets and all three classifiers.Figure 1 reports this privacy-utility relationship for LCED and MIMIC data.
- 3.3 Comparative analysis: Federated learning achieves comparable utility to centralized learning for both adverse drug reaction and mortality prediction.The comparison covers centralized learning, FL, and FL with ϵ-differential privacy.
- 3.3 Comparative analysis: Differential privacy provides a specified privacy level but causes significant deterioration in the federated model’s utility.The paper identifies this loss as especially relevant when federated sites are fewer than the very large numbers reported in prior work.
4 Conclusion
The paper implements federated learning for real-world health applications and examines differential privacy for protecting federated models. Its experiments show that differential privacy can significantly reduce healthcare model performance, motivating alternative privacy-preserving approaches.
- 4 Conclusion: The paper implements federated learning to address challenges associated with centralized learning for electronic health data.It also explores differential privacy for protecting federated learning models.
- 4 Conclusion: Experimental evaluation shows that differential privacy can cause a significant loss in model performance for healthcare applications.
- 4 Conclusion: The findings motivate research into alternative approaches for providing privacy in federated learning for healthcare applications.