Cybersecurity Policy Report, NIST Offers Guidance on ‘Privacy-Preserving Federated Learning’, (May 3, 2024)
The National Institute of Standards and Technology (NIST) yesterday published a blog on “privacy-preserving federated learning” (PPFL) that highlights “techniques for providing input privacy when data is vertically partitioned.”
The post explains that vertical partitioning is when training data is divided across parties with each holding different columns of data. This differs from horizontal partitioning as “training a model on vertically partitioned data is more challenging as it is generally not possible to train separate models on different columns of the data and then compose them afterwards,” it said.
Methods are needed to protect data when training on collective data, and one of the steps required for that is privacy-preserving entity alignment that matches corresponding records across different datasets, NIST said. The result of the entity alignment can be used to train a model in a similar fashion as in a horizontal partitioning scenario.
Two methods for privacy-preserving entity alignment include private set intersection (PSI) and Bloom filters. Privacy set intersection enables data linking between parties and reveals information only for rows of data that match a common key. The post cites to a pilot research project in which the PSI was applied to link the Department of Education with data from the National Student Loan Data System “to compute financial aid statistics without revealing students’ social security numbers.”
News: FederalLegislation DataPrivacy