Implementing Vertical Federated Learning Using Autoencoders: Practical Application, Generalizability, and Utility...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional machine learning (ML) requires centralized data, posing privacy and governance challenges.
• Federated learning (FL) offers a solution by training models across decentralized datasets.
• This study focuses on applying FL to vertically partitioned data.

Purpose of the Study:

• To implement FL on vertically partitioned data.
• To achieve ML model performance comparable to centralized approaches.
• To ensure raw data privacy is maintained during the process.

Main Methods:

• Vertical partitioning of three datasets (Adult income, Schwannoma, eICU).
• Training overcomplete autoencoder models at each site to generate latent data representations.
• Aggregating latent data for training a tabular neural network.
• Comparing FL model performance (accuracy, AUROC) against a centralized baseline.

Main Results:

• Autoencoder successfully transformed data into secure latent representations without domain knowledge.
• Transformed data exhibited different feature spaces and distributions, confirming data security.
• Minimal performance loss observed: accuracy loss ranged from 1.2% to 8.89%, AUROC loss from 0% to 1.12%.

Conclusions:

• An autoencoder-based ML model for vertically partitioned data was developed.
• The unsupervised approach eliminates the need for domain-specific knowledge at individual sites.
• This method provides a practical solution for building robust models when direct data sharing is not feasible, ensuring data protection.