Title

Obscure Logging: A Framework to Protect and Evaluate the Web Search Privacy

Abstract

Web Search Engine (WSE) is an inevitable software system used by people around the world to retrieve data from the web. WSE stores search queries to build the user’s profile and provides personalized results. These search queries hold identifiable information that can possibly compromise the privacy of the users. Preserving privacy in web search is the main concern of the users belonging to different walks of life. This research tries to highlight the loopholes imbedded in the available privacy preserving techniques. Besides, it aims at proposing some novel protocols with the least possible limitations. In this highly technological world, user is astonishingly surrounded by the amazingly advanced gadgets yet he is madly desirous to keep intact his privacy to the maximum. In order to preserve the Web search privacy of a user, this dissertation proposes a number of protocols such as a single group ObScure Logging (OSLo), a Multi Group ObScure Logging (MG-OSLo) and a Profile aware ObScure Logging (PaOSLo). This research work focuses on two main objectives. The first objective of this dissertation is to assess the local privacy and the profile privacy of a user through unlinkability and indistinguishability. The second objective of this dissertation is to evaluate the impact of group size, group count and the profile aware grouping on the local privacy and on the profile privacy of a user. Local privacy of proposed protocols has been evaluated by using probabilistic advantage being a curious entity and having linking query with the user. The profile privacy calculates the level of profile obfuscation using a privacy metric Profile Exposure Level (PEL). Computing the profile privacy of a user, a test has been performed over the same subset of AOL query log for two situations i.e. first, when the self-query submission is allowed and second, when self-query submission is not allowed. The privacy achieved by the proposed protocols has been compared with the state-of-the-art privacy-preserving protocol UUP(e) and co-utile protocol.

In the first protocol (OSLo), random users are grouped together to compute the impact of group size on the privacy of users in a single group design. In MG-OSLo, users are grouped by using non-overlapping group design and overlapping group design to measure the impact of group size and group count on the privacy of a user. The calculation depicts that the probability of linking query with the user depends on the group size and group count i.e. larger the group size or higher the group count lower the probability of linking query with the user. Whereas users, having dissimilar interest, are grouped together in PaOSLO, in order to evaluate the impact of profile aware grouping on the privacy. The results show that OSLo provides 9.37% better privacy as compared to the co-utile and 6.67% better privacy than UUP(e). The multi-group has a positive impact on the local privacy and on the profile privacy of a user. The MG-OSLo preserves 19.9% better privacy as compared to co-utile and 9.1% better as compared to UUP(e). Similarly, The profile aware grouping (PaOSLo) further improves the profile privacy as compared to UUP(e) and OSLo. The PaOSLo has 10% less PEL as compared to UUP(e) and 2.5% less as compared to OSLo when it is simulated on the same dataset.

Download full paper