Detecting Cascades from Weak Signatures

Inspired by cyber-security applications, we consider the problem of detecting an infection process in a network when the indication that any particular node is infected is extremely noisy. Instead of waiting for a single node to provide sufficient evidence that it is indeed infected, we take advantage of the graph structure to detect cascades of weak indications of failures. We view the detection problem as a hypothesis testing problem, devise a new inference algorithm, and analyze its false positive and false negative errors in the high noise regime. Extensive simulations show that our algorithm is able to obtain low errors in the high noise regime by taking advantage of cascading topology analysis.