Support Centre

You have out of 5 free articles left for the month

Signup for a trial to access unlimited content.

Start Trial

Continue reading on DataGuidance with:

Free Member

Limited Articles

Create an account to continue accessing select articles, resources, and guidance notes.

Free Trial

Unlimited Access

Start your free trial to access unlimited articles, resources, guidance notes, and workspaces.

International: RTA and NIST publish blog on input privacy for vertical federated learning

On May 2, 2024, the UK Department for Science, Innovation and Technology (DSIT) Responsible Technology Adoption Unit (RTA) in collaboration with the U.S. National Institute of Standards and Technology (NIST) published a blog post titled 'Input privacy for vertical federated learning.' In particular, the blog discusses techniques for providing input privacy when data is vertically partitioned.

Vertical partitioning

The blog begins by explaining vertical partitioning, where each participant in a federated learning project holds different parts of the data. This setup contrasts with horizontal partitioning, where each participant has the same kind of data but from different individuals. The blog emphasizes that training a model in a vertically partitioned system requires special techniques to protect privacy because data cannot be easily separated and recombined.

Key techniques for privacy

The blog introduces Private Set Intersection (PSI) as a technique to safely link data across different parties without exposing individual records. The blog explains that PSI protects privacy by varying how much information it reveals, from just the number of matching entries to more detailed data.

Another privacy technique discussed in the blog is Bloom filters, a tool that helps in identifying data matches without revealing the data itself. According to the blog, Bloom filters can sometimes lead to incorrect matches (false positives), which the blog points out as a form of automatic privacy protection, albeit with some risk of data leakage.

Balancing privacy and performance

The blog highlights the challenge of balancing efficient operation with the risk of leaking data. The Blog explains that while PSI and Bloom filters are effective for quick data matching, they can inadvertently reveal some information. The blog stresses the importance of carefully considering these risks when designing systems, mentioning that more secure methods exist but often at the cost of performance.

You can read the RTA blog post here and the NIST blog post here.