Neural Networks in Vulnerability Assessment

Network vulnerability assessment and analysis based on BP neural network algorithm

Because BP neural network algorithm is very sensitive to network structure, different network structures have different solving abilities. The more complex the neural network, the better its ability to deal with complex nonlinear problems, but the longer the training time. If the neural network structure is too simple, the network training is difficult to convergence or convergence time is too long. The topology of neural network includes the number of layers, the number of neurons per layer and the connections between neurons. In the BP network structure, the number of input neurons and output neurons is determined by the problem itself. Therefore, the design of BP network structure focuses on determining the number of hidden layer and the number of hidden layer neurons. The choice of the number of hidden layers depends on the complexity of the problem, and the relationship is shown in Fig. 4. The research shows that the increase in hidden layer can improve the ability of the network to solve complex nonlinear problems, but too much hidden layer can prolong the learning time of the network. For BP network, according to Kolmogorov theorem, three-layer BP network can complete arbitrary mapping from n dimension to m dimension, and implicit layer can meet the requirements. A hidden layer neural network, as long as the number of hidden layer neurons is reasonable, can meet the accuracy requirements. If the number of hidden layers changes from 1 to 2, it will not affect the accuracy much, but will make the network structure more complex and the training time will be greatly prolonged.

Fig. 4 Changes of hidden layer neurons and layer number under BP neural network

The representation and number of the input hidden layer neurons and the output hidden layer neurons of the neural network are determined by the problem itself and the requirements and representation of the data. Number of hidden layer neurons representation selection and parsing is a very complex mathematical problem, it and the requirement to the problem, the representation of a neural input layer and output unit type and the number is a direct relationship, often need to be based on the experience of the system designer and neural unit to determine, with the results of the experiment many times and therefore not may be an ideal implicit layer analytical formula for said. Because the expression and number of hidden layer neurons are too many, the learning time is too long, the error is not necessarily the least and the fault-tolerant and weak generalization ability may also be directly caused. Therefore, it is necessary to have an optimal number of hidden layer neurons.

According to the range of five grades of sensitivity index, 10 sets of data were randomly generated for each grade, and a total of 50 samples were formed. Choose one sample from each grade of sample (a total of five samples: samples 10, 20, 30, 40, and 50) as the test sample, the rest of the 45 samples as the training sample, which make all kinds of samples were distributed evenly, the BP neural network evaluation model is solved and set up without enough training samples and test samples. The sample processing of "self-coping ability" is similar, and the specific sample results are shown in Fig. 5.

Fig. 5 Sensitivity-level sample training results

Network vulnerability assessment and analysis based on SOM neural network algorithm

Analysis of computer network interface failure has generally four reasons of failure such as B1 interface problem, B2 network fault, B3 equipment existing congestion and B4 communication protocols, which are not compatible to as SOM neural network output node, MIB - 2 of 2 interface state of five signs, A1 interface problem, characteristic values of A2 type A3 output characteristic values of A4 network utilization and A5 unknown agreement rate as input nodes of networks; the fault training result is shown in Fig. 6.

Fig. 6 Training results of network fault samples

Since the SOM neural network is learning without teachers, the network will automatically cluster it. The number of network input vector elements is 5, ranging from [0, 1]. In order to improve the network mapping and achieve the best clustering effect, the competition layer of the network is designed as a 4 × 3 structure after multiple neural network training. The number of training steps affects the clustering performance of the network. Here, the training times are set to 100, and the results are shown in Fig. 7. The training function of neural network toolbox in MATLAB is used to train the SOM neural network. With the increase in training steps, the distribution of neurons is gradually reasonable. After the network training, the weights are fixed. After each input value, the network will automatically cluster it.

Fig. 7 Sample training results under SOM network

For the above after the weighting of training samples, the trained BP neural network is used in the simulation; for general need of simulation data samples, the neural network training time for each training sample P two updates, the original R the elements of the columns Q transformation into the column vector R + 1, thus increasing the one-dimensional elements, make the neural network input node number R + 1. In the simulation with the trained neural network, one dimension should be added to each sample to be simulated. Since the clustering of the samples to be simulated could not be known in advance, the added one-dimensional data should be set to the same value 1 to reflect the same clustering characteristics. When setting the weight of the sample, the preset weight in the training is also used to improve the sample data.