Development of a Nonsensitive Template for a 2D Ring vs. Square Discrimination Task

Linear models based on the channelized Hotelling observer were applied to binary discrimination tasks for enhanced information security in arms-control treaty verification. The model applies a transforma- tion matrix to binned projection data, generating a number of channelized values that can be combined optimally to yield a decision-making test statistic. This model processes data event-by-event, with only the channelized values being updated before the detected particle’s data is purged from memory. This prevents aggregation of sensitive information. The channelizing matrix is usually optimized to maximize the separation of the test statistic distributions that arise from imaging the two sources in the task. We present a method that could enable the host and monitor to agree on a nonsensitive channelizing matrix. If the host is able to define which parameters of their object are sensitive and put a tolerance on those parameters, they can add in penalty terms into the channelizing matrix optimization routine to penalize out sensitive information. The returned channelized values and test statistic would be indistinguishable when any source with a parameter value within the declared sensitive range is measured. To test this method, two-dimensional sources were simulated with the GEANT4 toolkit and classification tasks were performed using MATLAB. A 20 cm diameter ring source and a 20 cm long hollowed-out square source were imaged by the fast-neutron imaging system designed by Oak Ridge National Laboratory and Sandia National Laboratories. The imaging system contains 40x40 1 cm2 liquid scintillator pixels with a plastic coded aperture. The size of the sources was treated as the sensitive parameter in these studies, with a tolerance of 4 cm put on the lengths. The model was evaluated using the area under the ROC curve. The channelizing matrix resulting from the optimization routine effectively differentiates the different source types while being unable to discriminate along the sensitive size parameter.