The present invention discloses an SOM neural network-based malicious user discrimination method in cognitive radio. The method of the present invention utilizes an SOM neural network to learn the distribution characteristics of an input energy matrix, effectively classifies the input according to the learning results, and comprises the step of firstly introducing a suspicious degree concept, wherein the suspicious degree is distributed according to the number of the secondary users of each type after every time training. In order to eliminate the defects of the conventional SOM neural network, the present invention further provides an average suspicious degree concept. The concrete steps comprise obtaining an energy matrix, utilizing an SOM neural network algorithm to train the energy matrix to obtain a classification matrix, calculating the suspicious degree of each secondary user, constructing an index matrix and repeating the training process, averaging the suspicious degrees obtained at each time, namely obtaining the average suspicious degree, utilizing the average suspicious degree to classify the secondary users, and identifying the malicious users or the normal users.