D. Vaknin Ben Porath

Cascading failures in isotropic and anisotropic spatial networks induced by localized attacks and overloads

In this paper we study the Motter-Lai model of cascading failures induced by overloads in both isotropic and anisotropic spatial networks, generated by placing nodes in a square lattice and using various distributions of link lengths and angles. Anisotropy has not been earlier considered in the Motter-Lai model and is a real feature that may affect the cascading failures. This could reflect the existence of a preferred direction in which a given attribute of the system manifests, such as power lines that follow a city built parallel to the coast. We show that the anisotropy causes a greater spread of damage along the preferential direction of links. We also identify the critical linear size, $l_c$, for a square shaped localized attack, which satisfies with high probability that above $l_c$ the cascading disrupts the giant component of functional nodes, while below $l_c$ the damage does not spread. We find that, for networks with any characteristic link length, their robustness decreases with the strength of the anisotropy. We show that the value of $l_c$ is finite and independent of the system size (for large systems), both for isotropic and anisotropic networks. Thus, in contrast to random attacks, where the critical fraction of nodes that survive the initial attack, $p_c$, is usually below 1, here $p_c = 1$. Note that the analogy to $p_c = 1$ is also found for localized attacks in interdependent spatial networks. Finally, we measure the final distribution of functional cluster sizes and find a power-law behavior, with exponents similar to regular percolation. This indicates that, after the cascade which destroys the giant component, the system is at a percolation critical point. Additionally, we observe a crossover in the value of the distribution exponent, from critical percolation in a two-dimensional lattice for strong spatial embedding, to mean-field percolation for weak embedding.