Different kinds of user-generated data are increasingly used to tailor and optimize, through Machine Learning, the operation of online services and infrastructures. This typically requires sharing data among different partners, often including private data of individuals or business confidential data. While this poses privacy issues, the current state-of-the-art solutions either impose strong assumptions on the usage scenario or drastically reduce the data quality. In this paper, we evaluate through a generic framework the trade-offs between the accuracy of Machine Learning tasks and the achieved privacy (measured as similarity) on the input data, discussing trends and ways forward.