A graph theoretic linkage attack on microdata in a metric space

Certain methods of analysis require the knowledge of the spatial distances between entities whose data are stored in a microdata table. For instance, such knowledge is necessary and sufficient to perform data mining tasks such as nearest neighbour searches or clustering. However, when inter-record distances are published in addition to the microdata for research purposes, the risk of identity disclosure has to be taken into consideration again. In order to tackle this problem, we introduce a flexible graph model for microdata in a metric space and propose a linkage attack based on realistic assumptions of a data snooper’s background knowledge. This attack is based on the idea of finding a maximum approximate common subgraph of two vertex-labelled and edge-weighted graphs. By adapting a standard argument from algorithmic graph theory to our setup, this task is transformed to the maximum clique detection problem in a corresponding product graph. Using a toy example and experimental results on simulated data show that publishing even approximate distances could increase the risk of identity disclosure unreasonably.