Vickrey-Clarke-Groves for privacy-preserving collaborative classification

The combination of game theory and data mining opens new directions and opportunities for developing novel methods for extraction of knowledge among multiple collaborative agents. This paper extends on this combination, and motivated by the work of Nix and Kantarcioglu employs the Vickrey-Clarke-Groves (VCG) mechanism to achieve privacy-preserving collaborative classification. Specifically, in addition to encouraging multiple agents to share data truthfully, we facilitate preservation of privacy. In our model, privacy is accomplished by allowing the parties to supply a controlled amount of perturbed data, instead of randomised data, so long as this perturbation does not harm the overall result of classification. The critical point which determines when this perturbation is harmful is given by the VCG mechanism. Our experiment on real data confirms the potential of the theoretical model, in the sense that VCG mechanism can balance the tradeoff between privacy preservation and good data mining results.