Detecting and Localizing Adversarial Nodes Using Neural Networks

This work proposes a new method for securing the gossip algorithm for average consensus on communication networks. The gossip algorithm is appealing for its ability to harness distributed computational resources while adapting to arbitrarily connected networks without coordination overhead, however it is inherently vulnerable to the insider attack by adversarial node since each node locally updates its local states and passes information to its neighbors without supervision. In light of this, this work proposes new methods for detecting and localizing adversarial nodes using a neural network system. We show that our neural network-based method delivers a significantly improved detection and localization performance, compared to the state of the art.