Abstract: A method and a device for detecting risky types of data structures of a computer program code with a neural network. The neural network comprises neurons being related to each other by a topological arrangement involving a neighborhood definition. The neurons each comprising a vector for representing values of an input data space, at least one neuron having an associated label indicating the type of the neuron, and the data structures being detected comprising at least two elements. Information is extracted from at least two data elements from said at least one data structure and at least two input vectors are formed from said extracted information, the vectors being compatible with the vectors of the neurons. Said input vectors are compared with said vectors of the neurons, and the type of said basic data element is detected by using an associated label on the basis of said comparison.