4. Privacy-Preserving Embeddings

In this chapter, we illustrate main results on privacy-preserving embeddings. Here, security properties of embeddings are analyzed by considering two possible scenarios for their use. In the first case, a client submits a query containing sensitive information to a server, which should respond to the query without gaining access to the private information. This is discussed describing an authentication system in which a client submit an embedding of a physical characteristic of a device, and a verification server is able to match the embedding without revealing the actual physical characteristic. Interestingly, in this case the security properties of the embedding permit to combine it with existing biometric template mechanisms, enhancing the security of the system. In the second case, a large amount of sensitive data is stored in the cloud and a user should be able to make specific queries to the cloud without gaining access to the data. Here, we describe a universal embedding that preserves distances only locally. If data are stored in the cloud using this embedding, a user is able to retrieve data close to the query, but the complete geometry of the dataset remains hidden by the embedding and data cannot be recovered.