Bayesian Multiple Hypothesis Testing for Distributed Detection in Sensor Networks

In this paper we present a method for multiple hypothesis testing in sensor networks. Standard multiple testing methods ignore spatial dependencies between sensors, and thus suffer from a drastic loss of detection sensitivity. We introduce a Bayesian approach for taking the underlying spatial structure into account. By assigning a Gaus-sian process prior to a latent variable over the field, we locate areas where most of the received test statistics are caused by true signals. In these regions we relax the significance threshold in a manner which improves detection sensitivity while controlling overall false discovery rate at a tolerated level. The proposed method requires only minimal assumptions, but allows the user to incorporate their possible prior knowledge in to the model to refine inference. The benefits of the proposed method are demonstrated in simulation by comparing to standard multiple testing methods.