Data Protection in Big Data Analysis

"Big data" applications are collecting data from various aspects of our lives more and more every day. This fast transition has surpassed the development pace of data protection techniques and has resulted in innumerable data breaches and privacy violations. To prevent that, it is important to ensure the data is protected while at rest, in transit, in use, as well as during computation or dispersal. We investigate data protection issues in big data analysis in this thesis. We address a security or privacy concern in each phase of the data science pipeline. These phases are: i) data cleaning and preparation, ii) data management, iii) data modelling and analysis, and iv) data dissemination and visualization. In each of our contributions, we either address an existing problem and propose a resolving design (Chapters 2 and 4), or evaluate a current solution for a problem and analyze whether it meets the expected security/privacy goal (Chapters 3 and 5). Starting with privacy in data preparation, we investigate providing privacy in query analysis leveraging differential privacy techniques. We consider contextual outlier analysis and identify challenging queries that require releasing direct information about members of the dataset. We define a new sampling mechanism that allows releasing this information in a differentially private manner. Our second contribution is in the data modelling and analysis phase. We investigate the effect of data properties and application requirements on the successful implementation of privacy techniques. We in particular investigate the effects of data correlation on data protection guarantees of differential privacy. Our third contribution in this thesis is in the data management phase. The problem is to efficiently protecting the data that is outsourced to a database management system (DBMS) provider while still allowing join operation. We provide an encryption method to minimize the leakage and to guarantee confidentiality for the data efficiently. Our last contribution is in the data dissemination phase. We inspect the ownership/contract protection for the prediction models trained on the data. We evaluate the backdoor-based watermarking in deep neural networks which is an important and recent line of the work in model ownership/contract protection.