Enhancing Random Projection with Independent and Cumulative Additive Noise for Privacy-Preserving Data Stream Mining

Abstract The sensitive nature of many data streams necessitates data mining techniques that are privacy-preserving. This paper proposes two data perturbation methods for privacy-preserving stream mining based on a combination of random projection, random translation, and two alternative forms of additive noise: noise generated independently for each record and noise that accumulates over the lifetime of a stream. Variations of the known input-output Maximum A Posteriori (MAP) attack that can account for the combinations of perturbation techniques are proposed as a means of evaluating the privacy guarantees of the proposed perturbation methods. The capabilities of the proposed methods to resist privacy-breaching recovery attacks and retain accuracy in models trained on perturbed data are experimentally evaluated. Experimentation revealed that the cumulative noise injection scheme outperformed other schemes by achieving a superior trade-off between privacy and classification.