Compute-to-data is a method of using sensitive data without moving it. Instead of copying a dataset to the party that wants to analyse it, the analyst's algorithm is sent to the environment where the data is held, runs there under the data owner's control, and only the results, such as a trained model or summary statistics, are returned. The owner can approve which algorithms may run and keep the raw records private.
The approach is used in data marketplaces and for training machine learning models on data such as health or financial records, where data privacy rules or commercial interests prevent sharing. The term is closely associated with Ocean Protocol, a blockchain-based data marketplace that offers it as a feature. It belongs to the wider family of privacy-enhancing technologies, alongside secure multi-party computation and homomorphic encryption. Its main limit is that results can still leak information about the underlying data if queries are not controlled.
Ref 1300000287