The article deals with the issue of training a neural network when building algorithms for detecting violators in the decision-making block of modern detection tools. The feature of training neural networks in modern detection tools is to change the loss function under consideration, which takes into account the possible damage from the implementation of errors of the first and second kind in alarm systems. Based on the criterion of minimum average risk, it is advisable to minimize the probability of a false alarm (error of the first kind) with a fixed value of the probability of missing the target (error of the second kind). A new expression is obtained for updating the weights of the neural network during training, based on minimizing the new loss function. The process of training a neural network on a representative dataset of calculated information signal realizations and interference modeling is shown on the example of distributed magnetometric systems. It is proved that the recurrent neural network has high characteristics of detecting violators: for a given value of correct detection of 0,95, the probability of a false alarm was 5,9∙10-4.