Showing papers by "Charles G. Boncelet published in 2012"
21 Mar 2012
TL;DR: It is shown that a test on the sum of the reciprocal run lengths in a binary sequence typically performs as well as the classical Wald-Wolfovitz test, and significantly better in some cases.
Abstract: We consider an anomaly detection problem. We are interested in whether or not a stream of data contains an unusual number or distribution of positives. Abstractly, the problem can be stated as follows: given a binary string, we wish to determine if the number or distribution of 1's differs significantly from a known spontaneous rate. Furthermore, we consider the presence of an adversary who may try to distribute the 1's into ‘clusters’ to fool our test. We compare tests to detect this type of clustering to a simple test on the number of 1's, and show that clustered data is significantly easier to detect than i.i.d. data. We show that a test on the sum of the reciprocal run lengths in a binary sequence typically performs as well as the classical Wald-Wolfovitz test, and significantly better in some cases. We also show that if the length of the input stream is small, a simple additive correction term improves the detection rate of this test by a modest 1–2%.