Private Multi-File Retrieval From Distributed Databases

Suppose there are $N$ distributed databases each storing a full set of $M$ independent files. A user wants to retrieve $r$ out of the $M$ files without revealing the identity of the $r$ files. When $r=1$ it is the classic problem of private information retrieval (PIR). In this paper we study the problem of private multi-file retrieval (PMFR) which covers the case of general $r$. We first prove an upper bound on the capacity of PMFR schemes which indicates the minimum possible download size per unit of retrieved files. Then we design a general PMFR scheme which happens to attain the upper bound when $r\ge \frac{M}{2}$, thus achieving the optimal communication cost. As $r$ goes down we show the trivial approach of executing $r$ independent PIR instances achieves the near optimal communication cost. Comparing with the capacity-achieving PIR schemes, our PMFR scheme reduces the number of subpackages needed for each file from $N^M$ to $N^2$, which implies a great reduction of implementation complexity.