Non-metric multi-dimensional scaling for distance-based privacy-preserving data mining

Faculty of Science School of Computing Sciences Doctor of Philosophy by Khaled S. Alotaibi Recent advances in the field of data mining have led to major concerns about privacy. Sharing data with external parties for analysis puts private information at risk. The original data are often perturbed before external release to protect private information. However, data perturbation can decrease the utility of the output. A good perturbation technique requires balance between privacy and utility. This study proposes a new method for data perturbation in the context of distance-based data mining. We propose the use of non-metric multi-dimensional scaling (MDS) as a suitable technique to perturb data that are intended for distance-based data mining. The basic premise of this approach is to transform the original data into a lower dimensional space and generate new data that protect private details while maintaining good utility for distance-based data mining analysis. We investigate the extent the perturbed data are able to preserve useful statistics for distance-based analysis and to provide protection against malicious attacks. We demonstrate that our method provides an adequate alternative to data randomisation approaches and other dimensionality reduction approaches. Testing is conducted on a wide range of benchmarked datasets and against some existing perturbation methods. The results confirm that our method has very good overall performance, is competitive with other techniques, and produces clustering and classification results at least as good, and in some cases better, than the results obtained from the original data.