Seminar: Deep Learning for Network Attack and Defense

Student: Jordan Lanctot

Supervisor: Dr. Sean Cornelius

Abstract

Networked systems are vulnerable to attacks that remove one or a few key nodes, causing the connectivity to collapse. Through the use of Neural Network (NN) embedding frameworks,[1] Deep Leaning (DL) can be used to discover vulnerable nodes that will cause network failure.[2] This capacity to discover key features of networks of varying sizes and degrees, brings into question the security and robustness of infrastructure and networked systems. Can the vulnerabilities of real networks be obfuscated through partial network concealment strategies, allowing for a defensive response?

Here, we explore the ability of deep reinforcement learning to either: select nodes from a concealed graph to efficiently dismantle the network (network attack), or to thwart such an attacker by concealing a subset of the network’s links (network defense). We explore not only the capacity of these agents to learn against fixed heuristics, but also against each other. We show that for all of the heuristic approaches explored, agents are able to attack, or defend, a given network efficiently. When the two DL agents are pitted against one another, we find that, surprisingly, the defender has a natural advantage. This advantage results in tandem strategies which favour the defender over the span of many possible network configurations.