Logic of deception in machine learning

The issues of potential change in the work of artificial neural networks under various influences on training data is the urgent task. Violation of the correct operation of the artificial neural network with hostile effects on the training sample was called poisoning. The paper provides the simplest model of neural network formation in which the features used in training are based only on the predominance of the number of homogeneous elements. Changes in the samples of the training sample allow one to build Back Doors which, in turn, allow one to implement incorrect classification as well as embed errors into the software system, up to malicious code. The correct model of training sample poisoning which allows one to implement Back Door and triggers for classification errors is constructed in the paper. The simplest nature of the constructed model of functioning and formation of deception allows one to believe that the causal logic of the realization of a possible real attack on a complex artificial intelligence system has been restored correctly. This conclusion allows one in the future to correctly build the subsystems of monitoring, anomaly analysis, and control of the functionality of the entire artificial intelligence system.