Privacy preserving recommender systems

Badsha, S 2018, Privacy preserving recommender systems, Doctor of Philosophy (PhD), Science, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Badsha.pdf Thesis application/pdf 23.46MB
Title Privacy preserving recommender systems
Author(s) Badsha, S
Year 2018
Abstract The recommender systems help users find suitable and interesting products and contents from the huge amount of information that are available in the internet. There are various types of recommender systems available which have been providing recommendation services to users. For example Collaborative Filtering (CF) based recommendations, Content based (CB) recommendations, context aware recommendations and so on. Despite the fact that these recommender systems are very useful to solve the information overload problem by filtering interesting information, they suffer from huge privacy issues. In order to generate user personalized recommendations, the recommendation service providers need to acquire the information related to attributes, preferences, experiences as well as demands, which are related to users' confidential information. Usually the more information available to the service providers, the more accurate recommendations can be generated. However, the service providers are not always trustworthy to share personal information for recommendation purposes since they may cause serious privacy threats to users' privacy by leaking them to other parties or providing false recommendations. Therefore the user information must be protected prior to share them to any third party service provider to ensure the privacy of users.

To overcome the privacy issues of recommender systems several techniques have been proposed which can be categorized into decentralization, randomization and secure computations based approaches. In decentralization based approach, the central service providers are removed and the main controls of recommendation services are given to participant users. The main issue with this kind of approach is that to generate recommendations, the users need to be dependant to other users' availability in online services. If any user becomes offline, her information can not be used in the system. The randomization based techniques add noises to users data to obfuscate them from learning the true information. However the main issue is that adding noise affects recommendation accuracy. On the contrary, the secure computations preserve user information while providing accurate recommendations.

In this thesis we preserve user privacy by means of encrypting user information, specifically their ratings and other related information using homomorphic encryption based techniques to provide recommendations based on the encrypted data. The main advantage of homomorphic encryption based technique is that it is semantically secure and computationally it is hard to distinguish the true information from the given ciphertext. Using the homomorphic based encryption tools and techniques we build different privacy preserving protocols for different types of recommendation approaches by analyzing their privacy requirements and challenges. More specifically, we focus on different key recommendation techniques and differentiate them into centralized and partitioned dataset based recommendation techniques. From available recommendation techniques, we found that some of the existing and popular recommendation techniques like user based recommendation, item based recommendation and context aware recommendation can be grouped into centralized recommendation approach. In partitioned dataset based recommendation, the user information can be partitioned into different organizations and these organizations can collaborate with each other by gathering sufficient information in order to provide accurate recommendations without revealing their own confidential information. After categorizing the recommendation techniques we analyze the problems and requirements in terms of privacy preservation. Then for each type of recommendation approach, we develop the privacy preserving protocols to generate recommendations taking their specific privacy requirements and challenges into consideration.

We also investigate the problems and limitations of existing privacy preserving recommendations and found that the current solutions suffer from huge computation and communication overhead as well as privacy of users. In the thesis we identify the related problems and solve the issues using our proposed privacy preserving protocols. As an overall idea, our proposed recommendation protocols work as follows. The users encrypt their ratings using homomorphic encryption and send them to service providers. We assume the service providers are semi honest but curious, they follow the protocol but at the same time try to find new information from the available data. The service provider has the ability to perform homomorphic operations and it performs certain computations over encrypted data without learning any true information and returns the results to the query users who ask for recommendations. The system models of our privacy preserving protocols for different recommendation techniques differ from each other because of their different privacy requirements.

The proposed privacy preserving protocols are tested on various real world datasets. Based on the application areas of different recommendation approaches our gathered datasets are also different such as movie rating, social network, checkin information for different locations and quality of service of web services. For each proposed privacy preserving protocols we also present the privacy analysis and describe how the system can perform the computations without leaking the private information of users. The experimental and privacy analysis of our proposed privacy preserving protocols for different types of recommendation techniques show that they are private as well as practical.
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Science
Subjects Information Systems Management
Keyword(s) Cryptography
Recommender systems
Collaborative filtering
Encryption
Content based filtering
Ratings
Users
Versions
Version Filter Type
Access Statistics: 48 Abstract Views, 123 File Downloads  -  Detailed Statistics
Created: Tue, 05 Feb 2019, 12:58:59 EST by Keely Chapman
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us