This talk presents an improvement to the popular privacy-preserving data mining classifier induction technique, in which training data is perturbed prior to learning. We show how the seminal additive noise approach, proposed by Agrawal and Srikant in 2000, can be improved by using classifier ensembles. Particularly, for multiclass problems by using a classifier ensemble such as the Error-correcting Output Codes (and its extension to margin-based learners.) The method we propose improves significantly the quality of the learned classifier, while maintaining the same privacy levels as the original single classifier approach.