With ever increasing capacity for collecting, storing, and processing of data, there is also a high demand for intelligent knowledge discovery and data analysis methods. While there have been impressive advances in machine learning and similar domains in recent years, this also gives rise to concerns regarding the protection of personal and otherwise sensitive data, especially if it is to be analysed by third parties, e.g. in collaborative settings, where it shall be exchanged for the benefit of training more powerful models. One scenario is anomaly detection, which aims at identifying rare items, events or observations, differing from the majority of the data. Such anomalous items, also referred to as outliers, often correspond to problematic cases, e.g. bank fraud, rare medical diseases, or intrusions, e.g. attacks on IT systems.
Besides anonymisation, which becomes difficult to achieve especially with high dimensional data, one approach for privacy-preserving data mining lies in the usage of synthetic data. Synthetic data comes with the promise of protecting the users’ data and producing analysis results close to those achieved by using real data. However, since most synthetisation methods aim at preserving rather global properties and not characteristics of individual records to protect sensitive data, this form of data might be inadequate due to a lack of realistic outliers.
In this paper, we therefore analyse a number of different approaches for creating synthetic data. We study the utility of the created datasets for anomaly detection in supervised, semi-supervised and unsupervised settings, and compare it to the baseline of the original data.
