On May 2, 2024, the National Institute of Standards and Technology (NIST) published a blog on privacy-preserving federated learning (PPFL). In particular, NIST highlighted that vertical partitioning is where training data is divided across parties such that each party holds different columns of data, representing the challenge of being able to train on collective data while protecting the privacy of that data. Specifically, matching corresponding records without revealing the records themselves is a challenge.

What are the solutions to training separate models on different columns of data?

NIST recommended private set intersection (PSI) as a technique enabling data linking between parties, which reveals information to the participants only for rows of data that match on a common key.

NIST also recommended bloom filters, which use a collection of hash functions to enable efficient storage and lookup elements. Bloom filters are noted to protect privacy as there is no direct data sharing between different parties, only information regarding the potential absence of a specific feature, not the actual content of the data itself.

Notably, NIST clarified that the techniques above should be implemented in consideration of the level of risk faced by an organization and that additional techniques, such as homomorphic encryption, may be required.

You can read the blog post here.