Semi-local Time sensitive Anonymization of Clinical Data

A method for the anonymization of time-continuous data, which preserves the relation between the time- and value dimension is proposed in this work. The approach protects against linking- and distribution attacks by providing k-anonymity and t-closeness. Distributions can be generated from given sets using Distribution Clustering, according to the similarity of the curves, which serve as a replacement for the population distribution. Before the data is anonymized, it is split along the time-axis using Windowed Fréchet Splitting, to reduce the duration and information loss. The proposed approach employs bucketization using the Fréchet distance with an implicit maximum cost and implied t for closeness and multiple redistribution phases. The information loss, median relative error and achieved t for the closeness is low, and the runtime was reduced with the introduction of semi-local decisions.