A pattern recognition framework for the prediction of network vulnerabilities

Principles for the development of a toolbox aiming at detecting and predicting the vulnerabilities of complex networks are proposed in this article. They rest on pattern recognitions techniques, which leverage structural, dynamical, functional, and complexity features, which are selected to increase the sensitivity of the classifier to the detection of potential vulnerabilities in abnormal situations. Such an approach is expected to yield fast vulnerability prediction when compared with simulation using first-principle-based model of network. Furthermore, by leveraging actual data, pattern recognition techniques may circumvent the difficulty of obtaining accurate models in the absence of first principles (e.g., social networks). Vulnerability assessment is based on a risk function computed over all local representations of the network. Feature and representation extraction is then envisaged to seek a Pareto-efficient solution to the performance-robustness trade-off. Future directions for the development of the toolbox are discussed.