Accelerating Multi-Server Private Information Retrieval using GPUs

Master Thesis


Private Information Retrieval (PIR) allows a client to privately access an entry in a public database held by a server without leaking information about the client’s query. Example applications for PIR are private messaging apps [2] and privacy-friendly lookups in medical or patent databases. Modern multi-server PIR schemes outperform single server PIR in online runtime by reducing the amount of data each server has to touch during a request, but require that a subset of the servers are non-colluding. While great effort has been put into optimizing the online runtime, the storage requirements are expensive. A promising approach to reduce the storage for a single server is a centralized, powerful computer cluster that is optimized for offline runtime and distributes the necessary data among the servers.


The goal of this thesis is to integrate a centralized server in a PIR system based on RAID-PIR [1, 2]. This server shall provide the offline computation for the much less powerful PIR servers. At the end, the effectiveness of the centralized solution shall be benchmarked.


Firstly, a detailed literature study shall be conducted to gain a detailed knowledge of state of the art PIR schemes. The second step is a highly parallelized and optimized CUDA implementation to minimize the offline computation. Afterwards, the implementation shall be deployed and measured on a GPU cluster such as the Nvidia DGX-2 and/or a GPU node of the Lichtenberg High Performance Computer at TU Darmstadt. Finally, a comparison of the efficiency of the implemented system compared to previous work should be conducted by benchmarking runtimes and communication.


  • Good programming skills in C/C++
  • Basic knowledge of parallelization and CUDA are desirable
  • High motivation + ability to work independently
  • Knowledge of the English language, Git, LaTeX, etc. goes without saying


[1] Daniel Demmler, Amir Herzberg, and Thomas Schneider. RAID-PIR: Practical multi-server PIR. In 6. ACM Cloud Computing Security Workshop (CCSW’14), pages 45–56. ACM, 2014. PDF. (opens in new tab)

[2] Daniel Demmler, Marco Holz, and Thomas Schneider. OnionPIR: Effective protection of sensitive metadata in online communication networks. In 15. International Conference on Applied Cryptography and Network Security (ACNS’17), pages 599–619. Springer, 2017. PDF. (opens in new tab)