DO NOT DISTURB? Classifier Behavior on Perturbed Datasets

Exponential trends in data generation are presenting today’s organizations, economies and governments with challenges never encountered before, especially in the field of privacy and data security. One crucial trade-off regulators are facing regards the simultaneous need for publishing personal information for the sake of statistical analysis and Machine Learning in order to increase quality levels in areas like medical services, while at the same time protecting the identity of individuals. A key European measure will be the introduction of the General Data Protection Regulation (GDPR) in 2018, giving customers the ‘right to be forgotten’, i.e. having their data deleted on request. As this could lead to a competitive disadvantage for European companies, it is important to understand which effects deletion of significant data points has on the performance of ML techniques. In a previous paper we introduced a series of experiments applying different algorithms to a binary classification problem under anonymization as well as perturbation. In this paper we extend those experiments by multi-class classification and introduce outlier-removal as an additional scenario. While the results of our previous work were mostly in-line with our expectations, our current experiments revealed unexpected behavior over a range of different scenarios. A surprising conclusion of those experiments is the fact that classification on an anonymized dataset with outliers removed in beforehand can almost compete with classification on the original, un-anonymized dataset. This could soon lead to competitive Machine Learning pipelines on anonymized datasets for real-world usage in the marketplace.