Method, device and computer program for generating robust automatic learning systems and testing trained automatic learning systems

摘要

The present invention relates to a method of training the neural network 100, a method of testing the neural network 100, and also a method of detecting hostile examples that can deceive the neural network 100. An overlaid classification is propagated back through the second neural network 500, Maryland the output value of the second neural network 500 is used to determine whether the input of the neural network 100 is a hostile example. Established methods of the present invention are based on this usage of the second neural network 500. The present invention further relates to a computer program and an apparatus configured to perform one of the methods.