Deep Logarithmic Neural Network for Internet Intrusion Detection

Wang Zhendong,Xu Zhenyu,He Daojing,Chan Sammy
DOI: https://doi.org/10.1007/s00500-021-05987-9
IF: 3.732
2021-01-01
Soft Computing
Abstract:Intrusion detection is an important means to deal with network attacks. The traditional neural network needs to preprocess the sample data in intrusion detection, which may cause the loss in data accuracy, thus affecting the performance of intrusion detection. In this paper, a logarithm neuron (LOGN) is designed, and each LOGN is responsible for processing a kind of data feature. Different from the traditional neuron operation rules, the logarithm operation is carried out in the LOGN, and the base number of logarithm operations in each LOGN can be determined by the neural network according to the actual task requirements. Because the increase and decrease of logarithm operation are different under different logarithmic base values, the sensitivity of LOGN to the change of sample data feature values is also different, which makes LOGN have higher capability in data feature extraction. Based on LOGN, the logarithmic neural network layer (LOGL) is designed. Logarithmic neural network (LOGNN) is constructed by combining LOGLs with traditional deep neural network. To prevent the gradient of LOGNN from disappearing, a loss function of anti-gradient vanishing (AGLF) is proposed. Experimental results show that LOGNN has obvious advantages over traditional machine learning and deep neural network methods in terms of performance indices on NSL_ KDD and UNSW_ NB_15 data sets.
What problem does this paper attempt to address?