Resilient distributed optimization algorithm against adversary attacks
Resilient distributed optimization algorithm against adversary attacks
复制标题
DOI：
10.1109/icca.2017.8003106
复制
发表时间：
2017-07
期刊：
影响因子：
--
通讯作者：
Chengcheng Zhao;Jianping He;Qing-Guo Wang
中科院分区：
文献类型：
--
作者：
Chengcheng Zhao;Jianping He;Qing-Guo Wang
在分布式优化中，多个智能体的目标是最小化对应于一个决策变量的所有局部成本函数的平均值。近年来，针对攻击的分布式优化的弹性算法受到了一些关注，其中假设可容忍的攻击的最大数量受到网络连通性的严格限制。为了放宽这一假设，在本文中，我们提出了一个弹性分布优化算法，利用可信代理，这不能受到对手的攻击。我们证明了，所有正常代理的局部变量可以收敛在所提出的算法，如果可信代理诱导原网络的连通支配集。此外，我们利用正常代理的最终解将收敛到所有正常代理的局部函数的加权平均的凸最优集。我们还表明，可容忍的对手代理的数量是不受网络的连接所提出的算法。数值结果验证了该算法的有效性。
In the distributed optimization, multiple agents aim to minimize the average of all local cost functions corresponding to one decision variable. Recently, the resilient algorithms for distributed optimization against attacks have received some attention, where it is assumed that the maximum number of tolerable attacks is strictly limited by the network connectivity. To relax this assumption, in this paper, we propose a resilient distribution optimization algorithm by exploiting the trusted agents, which cannot be compromised by adversary attacks. We prove that local variables of all normal agents can converge under the proposed algorithm if the trusted agents induce the connected dominating set of the original network. Furthermore, we exploit that the final solution of normal agents will converge to the convex optima set of the weighted average of all normal agents' local functions. We also show that the amount of tolerable adversary agents is not limited by the network connectivity under the proposed algorithm. Numerical results demonstrate the effectiveness of the proposed algorithm.