Data Anonymization for Privacy Aware Machine Learning

摘要

The increase of data leaks, attacks, and other ransom-ware in the last few years have pointed out concerns about data security and privacy. All this has negatively affected the sharing and publication of data. To address these many limitations, innovative techniques are needed for protecting data. Especially, when used in machine learning based-data models. In this context, differential privacy is one of the most effective approaches to preserve privacy. However, the scope of differential privacy applications is very limited (e. g. numerical and structured data). Therefore, in this study, we aim to investigate the behavior of differential privacy applied to textual data and time series. The proposed approach was evaluated by comparing two Principal Component Analysis based differential privacy algorithms. The effectiveness was demonstrated through the application of three machine learning models to both anonymized and primary data. Their performances were thoroughly evaluated in terms of confidentiality, utility, scalability, and computational efficiency. The PPCA method provides a high anonymization quality at the expense of a high time-consuming, while the DPCA method preserves more utility and faster time computing. We show the possibility to combine a neural network text representation approach with differential privacy methods. We also highlighted that it is well within reach to anonymize real-world measurements data from satellites sensors for an anomaly detection task. We believe that our study will significantly motivate the use of differential privacy techniques, which can lead to more data sharing and privacy preserving.