Capacity Achieving Uncoded PIR Protocol based on Combinatorial Designs

TL;DR

This paper introduces a new capacity-achieving uncoded PIR protocol using combinatorial designs, significantly reducing subpacketization from exponential to linear in the number of servers, suitable for storage-constrained systems.

Contribution

The paper presents a novel uncoded PIR protocol based on combinatorial designs that achieves capacity with linear subpacketization, improving scalability for storage-constrained settings.

Findings

Achieves capacity with uncoded storage and fractional server storage.

Reduces subpacketization from exponential to linear in number of servers.

Uses system projection to multiple reduced systems for efficient retrieval.

Abstract

In this paper we study the problem of private information retrieval where a user seeks to retrieve one of the $F$ files from a cluster of $N$ non-colluding servers without revealing the identity of the requested file. In our setting the servers are storage constrained in that they can only store a fraction $\mu=t/N$ of each file. Furthermore, we assume that the files are stored in an uncoded fashion. The rate of a PIR protocol is defined as the ratio of the file size and the total number of bits downloaded. The maximum achievable rate is referred to as capacity. It was previously shown that there are capacity achieving PIR protocols when the file size is $N^F$ and complete files were stored on all the servers. These results were further extended for the case when servers store only a fraction of each file. However, the subpacketization $v$ of the files required is exponential in the number of servers $N$. We propose a novel uncoded PIR protocol based on combinatorial designs that are also capacity achieving when the file size is $v \times t^F$. Our protocol has linear subpacketization in the number of servers in contrast to previous work in storage constrained uncoded PIR schemes. In the proposed PIR protocol, the given system is projected to multiple instances of reduced systems with replicated servers having full storage capacity. The subfiles stored in these various instances are separately retrieved and lifted to solve the PIR problem for the original system.