準同型推論のための事前に暗号化されたデータの効率的なバッチ化

Abstract

Privacy-preserving homomorphic inference utilizes batch processing of encrypted data records, each of which has a private data portion on which inference is performed. Batching is made possible for a set of encrypted data records by techniques that ensure that each encrypted data record has its private data portion associated with a unique position relative to other data records. The set of encrypted data records are then summed to generate a single encrypted data record on which inference is performed. In a first embodiment, the private data portion of the subject is selectively uniquely positioned at run-time (where inference is applied). In a second embodiment, the private data portion of the subject is initially positioned together with stored data, preferably in an offline process. Then, at run-time, the individual encrypted data records are processed as necessary to adjust the private data portion to a unique position before batching.

Images