Towards an Inference Detection System Against Multi-database Attacks

Nowadays, users are permanently prompted to create web accounts when they buy online goods. This collected data gives an insight on the user, sometimes beyond the application scope. Inference attacks on databases represent an issue for data controllers when malicious processors attempt to guess sensitive data - to which they haven’t access - by inferring them using legally accessed data. Several inference attack detection systems address this problem in case of a single targeted database. But the issue remains unsolved in case of several databases to which the same users might have submitted their data. In this paper, we propose a global model and its associated graph representation named Global Instance Graph (GIG) representing the probabilistic and semantic dependencies inside each database, enriched by the dependencies between the different databases. The graph is obtained using privacy-preserving record linkage techniques and serves as a knowledge input to the inference attack detection system. We validate the GIG creation feasibility thanks to a proof of concept. Despite the quadratic creation time, the performances when data is queried from the databases are not affected since the GIG creation is performed offline.