Application Research Of Intrusion Detection Based On Improved GABP Neural Network

From early 90s of 20th century, with the rapid development of Internet, network security problems have become increasingly severe when computer networks give us infinite convenience. More and more network virus and hacker attack threat the security of network. Security and privacy of computer information have been seriously affected. Original static and passive defense technology of network has been unable to meet the higher safety requirements of the present network. Thus, a dynamic security defense technology --- Intrusion Detection technology is gradually becoming a key field within information security.Because the traditional rule-based intrusion detection technologies have problems, such as unmanageable rule base and tough-established statistical models, a idea of intrusion detection based on neural network has been put forward in recent years. Neural network has broad applications in the field of pattern recognition. In fact, network intrusion detection is pattern recognition pinpointing the network data flow and classify them as normal or abnormal data, so the use of fuzzy operation ability in neural network can solve certain problems of Intrusion Detection System. However, the traditional BP neural network easy to subside into the least value and slow convergence rate problems, at the same time, because of inherent limitations of the study methods, fundamentally improving the feed-forward back-propagation algorithm of neural network is very difficult. Through analyzing the theory and characteristics of traditional BP neural network, an improved genetic algorithm has been compounded to optimize the neural network weights. This method uses the strong global search capability and optimal method of genetic algorithm to overcome slow convergence rate and easy subsidy of least value of BP algorithm, the combination with the BP algorithm also solved the defect of can not find optimal solution in a short time when applying genetic algorithm alone. Using the gradient information of feed-forward back-propagation algorithm will avoid this situation. At the same time, dimensionally reducing neural network input data using principal component analysis. The application of principal component analysis can significantly reduce the dimensions of input data without losing original data information, improve the system's real-time property, and simplify the structure of the neural network. This paper, associated with matlab simulation, use mixed test data for various attacks and multiple test data for a single attack to experiment, finding that the improved genetic algorithm of BP neural network(GABP) has faster convergence rate , and this method is proved that it can converge to an overall optimal result. For different data sources, the tested recognition rate varies. For mixed test data for various attacks, GABP algorithm is markedly faster than BP algorithm for about 75%; for multiple test data for a single attack train, GABP algorithm can even achieve 99% recognition rate, the effect is great.