Cross Project Defect Prediction (CPDP) is a field of study where an organization lacking enough local data can use data from other organizations or projects for building defect predictors. Research in CPDP has shown challenges in using ``other'' data, therefore transfer defect learning has emerged to improve on the quality of CPDP results. With this new found success in CPDP, it is now increasingly important to focus on the privacy concerns of data owners.;To support CPDP, data must be shared. There are many privacy threats that inhibit data sharing. We focus on sensitive attribute disclosure threats or attacks, where an attacker seeks to associate a record(s) in a data set to its sensitive information. Solutions to this sharing problem comes from the field of Privacy Preserving Data Publishing (PPDP) which has emerged as a means to confuse the efforts of sensitive attribute disclosure attacks and therefore reduce privacy concerns. PPDP covers methods and tools used to disguise raw data for publishing. However, prior work warned that increasing data privacy decreases the efficacy of data mining on privatized data.;The goal of this research is to help encourage organizations and individuals to share their data publicly and/or with each other for research purposes and/or improving the quality of their software product through defect prediction. The contributions of this work allow three benefits for data owners willing to share privatized data: 1) that they are fully aware of the sensitive attribute disclosure risks involved so they can make an informed decision about what to share, 2) they are provided with the ability to privatize their data and have it remain useful, and 3) the ability to work with others to share their data based on what they learn from each others data. We call this private multiparty data sharing.;To achieve these benefits, this dissertation presents LACE (Large-scale Assurance of Confidentiality Environment). LACE incorporates a privacy metric called IPR (Increased Privacy Ratio) which calculates the risk of sensitive attribute disclosure of data through comparing results of queries (attacks) on the original data and a privatized version of that data. LACE also includes a privacy algorithm which uses intelligent instance selection to prune the data to as low as 10% of the original data (thus offering complete privacy to the other 90%). It then mutates the remaining data making it possible that over 70% of sensitive attribute disclosure attacks are unsuccessful. Finally, LACE can facilitate private multiparty data sharing via a unique leader-follower algorithm (developed for this dissertation). The algorithm allows data owners to serially build a privatized data set, by allowing them to only contribute data that are not already in the private cache. In this scenario, each data owner shares even less of their data, some as low as 2%.;The experiments of this thesis, lead to the following conclusion: at least for the defect data studied here, data can be minimized, privatized and shared without a significant degradation in utility. Specifically, in comparative studies with standard privacy models (k-anonymity and data swapping), applied to 10 open-source data sets and 3 proprietary data sets, LACE produces privatized data sets that are significantly smaller than the original data (as low as 2%). As a result LACE offers better protection against sensitive attribute disclosure attacks than other methods.
展开▼
