PrivacyGroup Event:2016/03/04 Privacy as secure multiparty communication
- As Hong mentioned, perturbation techniques mean there is a trusted third party who holds all private data. This approach addresses that weakness with a technique that allows a third party to compute statistics on a private data set without learning anything beyond those statistics. The statistics themselves can be used to infer private data, however, so it is significant that this approach can be composed with a perturbation method (e.g. Dwork's Laplace mechnanism) to provide better privacy. It is significant that each data subject locally computes part of the statistics -- this is a massively distributed arrangement.
- Formulates a privacy problem in terms of secure multiparty communication. Makes clear the threat model for the above article is passive, honest-but-curious.
- A hit-and-miss discussion is Cryptography and Cryptographic Protocols, by Goldreich
- Description of the definition itself is helpful, as is discussion of why DP algorithms are less likely to overfit to a hold out set