Nowadays organizations all over the world are dependent on mining gigantic datasets. These datasets typically contain delicate individual information, which inevitably gets exposed to different parties. Consequently privacy issues are constantly under the limelight and the public dissatisfaction may well threaten the exercise of data mining and all its benefits. It is thus of great importance to develop adequate security techniques for protecting confidentiality of individual values used for data mining. In the last 30 years several techniques have been proposed in the context of statistical databases. It was noticed early on that non-careful noise addition introduces biases to statistical parameters, including means, variances and covariances, and sophisticated techniques that avoid biases were developed. However, when these techniques are applied in the context of data mining, they do not appear to be bias-free. Wilson and Rosen (2002) suggest the existence of Type Data Mining (DM) bias that relates to the loss of underlying patters in the database and cannot be eliminated by preserving simple statistical parameters. In this paper we propose a noise addition framework specifically tailored towards the classification task in data mining. It builds upon some previous techniques that introduce noise to the class and the so-called innocent attributes. Our framework extends these techniques to the influential attributes; additionally, it caters for the preservation of the variances and covariances, along with patterns, thus making the perturbed dataset useful for both statistical and data mining purposes. Our preliminary experimental results indicate that data patterns are highly preserved suggesting the non-existence of DM bias.