A Multi-Convolutional Neural Network Framework For Network Connectivity Robustness Prediction

Connectivity robustness reflects the ability of a complex network maintaining its basic connected structures and functions under various external attacks. Conventionally, connectivity robustness is evaluated by simulating attacks, which may take expensive computational resources, especially for large-scale networked systems. Past studies have shown that convolution neural networks (CNN) have excellent performance in both regression and classification. Complex networks are transformed into grey images and processed by CNNs. In this paper, a multi-CNN framework is come out to predict the connectivity robustness, where each CNN performs either network classification or robustness prediction. Experimental studies are carried on four network topologies under random and targeted attacks. The simulations results shows that the proposed multi-CNN approach can precisely predict the connectivity robustness of complex networks. Further experiments also shows that multi-CNN performs better than single CNN method.