With the increasing interest in developing pervasive computing technologies there is growing recognition of the problems of maintaining user privacy. In the Daidalos pervasive system this is achieved primarily through the use of virtual identities, which are used to conceal the real identity of the user. One problem with this lies in determining to what extent the user should be engaged in the decisions relating to the selection of virtual identities, and what can be done automatically. The solution lies in creating a set of user preferences to assist in taking these decisions, refining them through the use of machine learning techniques. This paper outlines the approach being investigated and describes how this will be achieved when the processes involved in building up user preferences are not trusted.