Designing a Cost Function to Assess a Neural Network to Detect Distributed Denial of Service Attacks

This paper presents a model for designing a cost function for neural networks. The proposed procedure consists of enriching a basis cost function with distinguishable features as the coefficients to create a highly sensitive cost function. Since the Internet traffic data that contains distributed denial of service is not balanced, exaggerating the anomalous part of data leads to a better classification and data class separation. To develop the proposed cost function, sensitivity analysis is used as a measure to assess and test the cost function’s parameters. The principle components of the most sensitive cost function are extracted. The most efficient principle component with the highest variance is selected as the weights for the selected cost function. Therefore, the highest separation between two normal and anomalous clusters can be gained.