Federated machine learning with Anonymous Random Hybridization (FeARH) on medical records

Jianfei Cui, Dianbo LiuarXiv 2019 · OpenReview

TL;DR

This paper introduces FeARH, a novel federated learning approach that uses "anonymous random hybridization" to protect patient privacy in medical record analysis. By randomly exchanging model parameters between participants before aggregation, FeARH significantly reduces the risk of data leakage from untrustworthy parties, achieving comparable performance to traditional methods.

What Problem Does This Paper Address?

Training machine learning models on electronic medical records (EMRs) is crucial for advancements in healthcare, but it's hampered by privacy concerns and regulatory hurdles that make centralizing data difficult. Federated learning offers a solution by training models across distributed data silos without moving raw data. However, a key challenge remains: even in federated learning, intermediate parties (like a central analyzer) might be untrustworthy and could potentially infer sensitive patient information from the model parameters they handle. This paper tackles this vulnerability by proposing a method to further anonymize the parameter exchange process, making it harder for any single party to reconstruct private data.

Methods

The core innovation of this paper is the Federated machine learning with Anonymous Random Hybridization (FeARH) algorithm. Unlike traditional federated learning where models are trained locally and then aggregated, FeARH introduces a hybridization step. After local training, instead of directly sending models to a central analyzer, participants randomly exchange a portion of their model parameters with each other. This "hybridization" process, controlled by an exchange rate ($\gamma$), effectively breaks the direct link between a specific local dataset and the final aggregated model. The parameters are then aggregated by the central analyzer. This randomization aims to prevent an untrustworthy central analyzer from deducing sensitive information about individual patient data from the model parameters.

Datasets

The study utilizes the eICU Collaborative Research Database. This is a large, multi-center critical care database containing information on over 30,760 patients, including medicines taken and mortality. The data was split into training (70%), testing (20%), and validation (10%) sets. For experimental purposes, this data was further divided into 8 sub-datasets for each of the 8 simulated data owners.

Key Insights and Ideas

  • Privacy through Parameter Randomization: The fundamental idea is that by randomly shuffling and exchanging parts of model parameters before aggregation, the direct traceability of specific data points to model updates is obscured. This is a crucial step beyond standard federated learning, which still relies on a potentially untrustworthy central aggregator.
  • "Active" Adversary Mitigation: FeARH is designed to counter an "active" adversary, meaning a central party that can not only observe but also potentially manipulate the training process or model parameters. The hybridization step makes such manipulation less effective in revealing private information.
  • Trade-off between Privacy and Performance: The paper explores the impact of the hybrid exchange rate ($\gamma$). While higher rates offer better privacy by increasing randomization, they can lead to a slight decrease in model performance (AUCROC and AUCPR). This highlights a common trade-off in privacy-preserving machine learning.

Key Results

The paper compares FeARH with centralized learning and original federated learning on the eICU dataset using AUCROC and AUCPR metrics.

  • Neural Network Performance:
    • Centralized Learning: AUCROC of 0.8412, AUCPR of 0.6794.
  • Original Federated Learning: AUCROC of 0.8452, AUCPR of 0.7019.
  • FeARH: AUCROC of 0.8312, AUCPR of 0.6778. While FeARH shows slightly lower performance than the other two methods, the difference is marginal, suggesting it achieves comparable accuracy while enhancing privacy.

  • Logistic Regression Performance:

    • Centralized Learning: AUCROC of 0.8412, AUCPR of 0.4579.
  • Original Federated Learning: AUCROC of 0.7816, AUCPR of 0.5989.
  • FeARH: AUCROC of 0.7864, AUCPR of 0.6206. Interestingly, for logistic regression, FeARH slightly outperforms original federated learning in AUCPR.

  • Privacy Preservation: An additional experiment using a Generative Adversarial Network (GAN) as an attack model demonstrated that FeARH is significantly more resistant to information leakage compared to original federated learning. The attacker achieved much lower accuracy (0.6919 vs. 0.9957) and higher loss (3.0011 vs. 0.0976) when attacking FeARH.

Limitations and Future Work

  • Performance Degradation: The authors acknowledge that FeARH can lead to a slight decrease in model performance compared to centralized or original federated learning, especially with higher exchange rates. The optimal balance between privacy and performance needs to be carefully considered.
  • Exchange Rate Selection: The choice of the hybrid exchange rate ($\gamma$) is presented as a subjective parameter, and its impact on performance and privacy is explored but not definitively optimized. Future work could focus on adaptive methods for selecting this rate.
  • Generalizability: While the paper focuses on neural networks and logistic regression, it suggests that the proposed algorithm strategy could be useful for other machine learning models. Further experiments on different model architectures and datasets would be valuable.
  • Computational Overhead: The hybridization step adds computational complexity. While not explicitly detailed, this is a potential area for optimization.

Summary updated

Abstract

Sometimes electrical medical records are restricted and difficult to centralize for machine learning, which could only be trained in distributed manner that involved many institutions in the process. However, sometimes some institutions are likely to figure out the private data used for training certain models based on the parameters they obtained, which is a violation of privacy and certain regulations. Under those circumstances, we develop an algorithm, called 'federated machine learning with anonymous random hybridization'(abbreviated as 'FeARH'), using mainly hybridization algorithm to eliminate connections between medical record data and models' parameters, which avoid untrustworthy institutions from stealing patients' private medical records. Based on our experiment, our new algorithm has similar AUCROC and AUCPR result compared with machine learning in centralized manner and original federated machine learning, at the same time, our algorithm can greatly reduce data transfer size in comparison with original federated machine learning.
BibTeX
@misc{cui2019federated,
  author = {Jianfei Cui and Dianbo Liu},
  title = {{Federated machine learning with Anonymous Random Hybridization (FeARH) on medical records}},
  year = {2019},
  eprint = {2001.09751},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.09751},
}
Figure