The Grapevine Web: Analysing the Spread of False Information in Social Networks with Corrupted Sources

Jacques Bara, Charlie Pilgrim, Paolo Turrini, Stanislav Zhydkov

Abstract: We study the problem of noisy information propagation in networks, where a small number of sources send messages across the network and agents use Bayesian updates to make inferences about the state of the world from the received messages. We provide upper bounds on the total number of sources necessary for learning on a given network and refine the bound in the case of small-world networks. We then extend the model to include an adversarial attacker, who can corrupt some of the information sources. We find that there is an optimal greedy attacking strategy in the case of a single learner, while the multi-learner case is not always solved optimally using greedy approaches. However, despite the influence function not being submodular, we show that the greedy algorithm performs well in practice. We also show that much simpler heuristics, which only look at centrality measures, can also provide a good basis to calculate successful attacking strategies. Finally we analyse the loss of optimality in the case when the attacker has incomplete information about the network and has to estimate the influence of source corruption heuristically. We use real-world social networks, as well as random network models, to empirically evaluate the effectiveness of attacking strategies and suggest a variety of measures to counteract them.