Research On The Structural Optimization Of BP Neural Network Algorithm For Anomaly Detection

With the development of the Internet, its richful information brings great convenience to our life, but it also brings network security problems.Traditional network threat mainly from viruses, Trojan horses, password invasion, node attack attacked the traditional method, with the arrival of the era of mobile Internet, various kinds of new attack methods emerge in endlessly, network security situation is not optimistic. Flow detection, as an active defense security technology, provides real-time protection against internal attacks, external attacks and misuse of the system, to build the system following the second gates of the firewall. The current flow detection technology has been developed and integrated into neural network, ant colony algorithm and other intelligent technologies.The main research work of this paper is a method of detecting abnormal traffic. Abnormal traffic means deviation from the normal range of network traffic anomaly traffic flow characteristics reflected in the value fluctuations. Testing process:First, collect traffic data preprocessing, feature extraction method proposed PCA, to achieve linear dimensionality reduction; secondly using BP algorithm structure optimized for training learning; and finally, verification and generalization algorithms optimized detection rate.BP algorithm optimization stage, to improve the generalization performance algorithm is proposed to optimize the structure of the neural network direction, a direction is to optimize the input layer nodes can be used PCA dimension reduction, the other direction is to adjust the hidden layer, it can reduce the number of hidden layers and each hidden layer nodes. Optimized for hidden layer nodes, right out of the proposed connection weight adjustment, specifically the right to a non-zero value to zero pull, causing the error if it means significantly increased weight must be preserved, otherwise it is eliminated, ultimately makes the hidden layer connection weights of zero nodes removed, hidden layer to achieve structural optimization. Traditional BP neural network pruning method has limitations in the right to punish, we prove the convergence of the algorithm has eliminated the right that the algorithm is effective, and better than the right methods of punishment in generalization performance. Optimized BP neural network structure by more than two directions, effectively reduce the time complexity of the algorithm. In summary, this proposed network traffic optimization algorithm BP neural network anomaly detection method that the improved BP neural network algorithm to optimize and adjust the network structure, focus on reducing the complexity of the topologies BP, can effectively improve the algorithm generalization ability, enhance backbone network nodes in real-time anomaly detection. Experimental results show that the optimization algorithm time complexity and convergence properties are superior to traditional BP algorithm is applied to a network anomaly detection system in the detection rate and time efficiency is significantly superior to traditional BP algorithm.