Data Injection Attacks in Randomized Gossiping

The subject of this paper is the detection and mitigation of data injection attacks in randomized average consensus gossip algorithms. It is broadly known that the main advantages of randomized average consensus gossip are its fault tolerance and distributed nature. Unfortunately, the flat architecture of the algorithm also increases the attack surface for a data injection attack. Even though we cast our problem in the context of sensor network security, the attack scenario is identical to existing models for opinion dynamics (the so-called DeGroot model) with stubborn agents steering the opinions of the group toward a final state that is not the average of the network initial states. We specifically propose two novel strategies for detecting and locating attackers and study their detection and localization performance numerically and analytically. Our detection and localization methods are completely decentralized and, therefore, nodes can directly act on their conclusions and stop receiving information from nodes identified as attackers. As we show by simulation, the network can often recover in this fashion, leveraging the resilience of randomized gossiping to reduced network connectivity.