Quickest Anomaly Detection in Sensor Networks With Unlabeled Samples

The problem of quickest anomaly detection in networks with unlabeled samples is studied. At some unknown time, an emerges the network and changes data-generating distribution sensor. data vector received by fusion center at each time step undergoes arbitrary permutation its entries (unlabeled samples). goal to detect minimal delay subject false alarm constraints. With samples, existing approaches that combines local cumulative sum (CuSum) statistics cannot be used anymore. Several major questions include whether still possible without label information, if so, what fundamental limit how achieve that. Two cases static dynamic are investigated, where sensor affected may or not change time. For two cases, practical algorithms based on ideas mixture likelihood ratio and/or maximum estimate constructed. Their average delays rates theoretically characterized. Universal lower bounds for a given rate also derived, which further demonstrate asymptotic optimality algorithms.