Privacy Preserving Average Consensus with Different Privacy Guarantee

Abstract

Privacy preserving average consensus problem concerns protecting the initial states of participants not to be disclosed while achieving average consensus. In existing literatures, most of works consider providing privacy guarantee for agents when all the other participants are viewed as the same type of privacy attackers. However, for an agent, not all the other participants are untrustworthy. The agent can set weaker privacy demand against these credible participants to obtain potential utility improvement. In this paper, we consider that network is composed of several groups. Within each group, agents treat participants within and outside the group as two types of attackers. Then, a private average consensus algorithm (PACA) is proposed to provide different privacy protection against participants within and outside the group, by perturbing agents' initial states with random noises. Based on the definition of $\varepsilon$ -KL privacy, we prove that PACA preserves distinct degrees of privacy guarantee against agents within and outside the group. Moreover, the convergence accuracy is analyzed according to the deviation between the final value and the true average one. Finally, simulations are conducted to evaluate the performance of PACA.