Latent Semantic Indexing (LSI) Based Distributed System and Search On Encrypted Data

Abstract

Latent semantic indexing (LSI) was initially introduced to overcome the issues of synonymy and polysemy of the traditional vector space model (VSM). LSI, however, has challenges of its own, mainly scalability. Despite being introduced in 1990, there are few attempts that provide an efficient solution for LSI, most of the literature is focuses on LSI’s applications rather than improving the original algorithm. In this work we analyze the first framework to provide scalable implementation of LSI and report its performance on the distributed environment of RAAD. The possibility of adopting LSI in the field of searching over encrypted data is also investigated. The importance of that field is stemmed from the need for cloud computing as an effective computing paradigm that provides an affordable access to high computational power. Encryption is usually applied to prevent unauthorized access to the data (the host is assumed to be curious), however this limits accessibility to the data given that search over encryption is yet to catch with the latest techniques adopted by the Information Retrieval (IR) community. In this work we propose a system that uses LSI for indexing and free-query text for retrieving. The results show that the available LSI framework does scale on large datasets, however it had some limitations with respect to factors like dictionary size and memory limit. When replicating the exact settings of the baseline on RAAD, it performed relatively slower. This could be resulted by the fact that RAAD uses a distributed file system or because of network latency. The results also show that the proposed system for applying LSI on encrypted data retrieved documents in the same order as the baseline (unencrypted data).