On Preserving Privacy

Abstract

This thesis explores homomorphic encryption on queries to cloud stored documents. Moreover, the aim of this thesis is to explore the use of semantic search on encrypted word embeddings with end-to-end privacy. A concrete implementation of a secure semantic search application that stores documents in a database which allows for efficient retrieval (using Locality Sensitive Hashing) and computation of embedding similarities is presented. Experiments were conducted to benchmark the performance of homomorphic operations on encrypted data. CKKS was the homomorphic encryption scheme used in these experiments, because CKKS works with vectors of real numbers, which is what word embeddings are. The experiments focused on how much of the operations can be offloaded to the server and also the accuracy between decrypted ciphertexts and plaintexts after computations. Our results show that achieving high accuracy between decrypted ciphertexts and plaintexts does not decrease performance, however it does limit the level of security depending on the set parameters. We concluded that homomorphic encryption is feasible for our specific use-case and could potentially allow almost 300 000 similarity computations per second given a server cluster of 8 hosts each having an Nvidia 4090 GPU.