The present invention relates to a technology for detecting abnormal web traffic, which can autonomously learn characteristics of target data to be learned and predicted by a web traffic detection algorithm, thereby enabling more relevant characteristics to be extracted than a conventional method and enabling various attacks to be detected. According to an embodiment of the present invention, provided is a method for detecting abnormal input data by using a convolutional neural network, the method comprising the steps of: (a) when input data including one or more sequentially arranged pieces of text is acquired as a training set, converting, by an apparatus, the input data into data in a matrix form, or supporting the conversion; (b) performing, by the apparatus, a convolution operation on the data in the matrix form by using the predetermined number of first kernels, or supporting this performance; (c) converting, by the apparatus, the data, on which the convolution operation has been performed, into data in a predetermined matrix form, and performing a fully connected operation, adapted to generate a neural network layer, by using the data in the predetermined matrix form or supporting this performance; (d) performing, by the apparatus, a deconvolution operation, reverse to the convolution operation, on data in a matrix form generated as a result of the performance of the fully connected operation, or supporting this performance; and (e) calculating, by the apparatus, a value of a difference between the converted input data and a result of the deconvolution operation, or supporting the calculation.;COPYRIGHT KIPO 2016
展开▼
