This report overviews the preliminary design of new privacy preserving (PP) primitives for three main data analytics operations, namely: neural network classification and training, clustering and counting. The document focuses on these three particular algorithms as these are mainly derived from the definition of the PAPAYA use cases in Deliverable D2.1. All these newly proposed solutions aim at enabling an untrusted third-party data processor (the PAPAYA platform, in this case) to perform the underlying operations over protected data. Thanks to these primitives, data owners will be able to extract valuable information from their protected data while being cost-effective and accurate. This report first reviews the existing cryptographic tools including homomorphic encryption, secure multi-party computation, functional encryption and differential privacy, which are used as building blocks for the design of the newly proposed PAPAYA primitives. Then, the PAPAYA solutions are further described under the following three categories:
- Privacy preserving neural networks (PP-NN)
- Privacy preserving clustering
- Privacy preserving counting
See the full document in the link below.