Feature selection on high-dimensional networks plays an important role in understanding of biological mechanisms and disease pathologies. It has a broad range of applications. Recently, a Bayesian nonparametric mixture model (Zhao, Kang, and Yu 2014) has been successfully applied for selecting gene and gene sub-networks. We extend this method to a unified approach for feature selection on general high-dimensional networks; and we develop an R package, the Bayesian network feature finder (BANFF), providing a package of posterior inference, model comparison, and graphical illustration of model fitting. A parallel computing algorithm for the MCMC based posterior inference and an EM based algorithm for posterior approximation are added. Also, a double Metropolis-Hasting sampler (Liang 2010) is adopted for hyper-parameters searching, which is more efficient and accurate in determining the settings of hyper-parameter. Here we provide an instruction along with examples. Particularly, we demonstrate the use of BANFF on analyzing a protein-protein interaction network and a longitudinal Positron emission tomography image data in the Alzheimer's disease neuroimaging initiative (ADNI) study.