Visualizing Convolutional Neural Network Models’ Sensitivity to Nonnatural Data Order

Convolutional neural networks (CNN) have revolutionized image recognition technology and found applications in various nonimage-related fields. For nonnatural data, such as cybersecurity packets, in which the data sample order is not defined by nature, some models trained on certain orderings of data perform better than when trained with other orderings. Some orderings create patterns from which the CNN extracts better features. Understanding how to best order the training data to improve CNN performance is not well-studied. In this study, we investigate this problem by examining malware infections using different CNN models and include visualization tools to enhance our analysis. We define a functional algorithm for ordering and show that order importance in CNNs is model dependent. In addition, we show that depending on the model, statistical relationships are crucial in establishing order with better performance.