Detection of data injection attacks in decentralized learning

Gossip based optimization and learning are appealing methods that solve big data learning problems sharing computation and network resources when data are distributed. The main advantage these methods offer is that they are fault tolerant. Their flat architecture, however, expands the attack surface in the case of a data injection attack. We analyze the effects of data injection on the asymptotic behavior of the network and draw a parallel with the case of opinion dynamics in a network where zealots inject opinions to mislead a community. We further propose a possible decentralized detection of such attacks and analyze its performance.