Differentially Private Learning of Undirected Graphical Models using Collective Graphical Models