Scalable and Privacy-Preserving Federated Principal Component Analysis.

Area of Science:

• Data Science
• Cryptography
• Distributed Computing

Background:

• Principal Component Analysis (PCA) is crucial for dimensionality reduction.
• Federated learning presents challenges in maintaining data confidentiality during collaborative analysis.
• Existing methods for secure PCA are often inefficient or provide approximate results.

Purpose of the Study:

• To develop a secure and efficient federated Principal Component Analysis (PCA) system.
• To ensure data confidentiality for both original data and intermediate results in a distributed setting.
• To achieve accurate PCA results comparable to centralized methods.

Main Methods:

• SF-PCA system leverages multiparty homomorphic encryption, interactive protocols, and edge computing.
• It interleaves computations on local cleartext data with operations on encrypted data.
• The system operates under a passive-adversary model with up to all-but-one colluding parties.

Main Results:

• SF-PCA achieves accuracy comparable to non-secure centralized PCA, irrespective of data distribution.
• The system demonstrates linear or better scalability with dataset dimensions and number of data providers.
• SF-PCA is significantly faster (3x-250x) than existing privacy-preserving PCA alternatives.

Conclusions:

• SF-PCA provides a practical and efficient solution for secure federated PCA on private, distributed datasets.
• The system offers superior precision and performance compared to current approaches.
• This work highlights the feasibility of applying advanced cryptographic techniques for confidential data analysis.