Today’s society is witnessing not only an evergrowing dependency on data, but also an increasingly pervasiveness of related analytics and machine learning applications. From business to leisure, the availability of services providing answers to questions brings great benefits in diverse domains. On the other side of the coin, the need to provide input data that the services need to compute a response. However, some data may be considered sensitive or confidential and users would legitimately be reluctant to release them to third parties. Considering classification tasks in machine learning applications, we introduce our PriSM (Privacy-friendly Support vector Machine) approach for computing a privacy-friendly model. PriSM anticipates the training phase of the classifier with a phase for discovering correlations among attributes that can indirectly expose sensitive information. It then trains the classifier excluding from consideration not only sensitive attributes but also other sets of attributes that have been learned as correlated to them. The result is a privacy-friendly classifier that does not require any of such information as input from the users. Our experimental evaluation on both synthetic and real-world datasets confirms the effectiveness of PriSM in protecting privacy while maintaining classification accuracy.