In today?s information society, privacy protection has become a very important concern. In this paper we research the inference problems due to functional dependencies (FD) and multi-valued dependencies (MVD) in multilevel secure database (MLS) with element classification instances. To deal with the secure problem brought by inference channels we present our FD and MVD based inference control algorithms working on the finest-grained data level which greatly improve the availability of data and minimize the information loss.