Differentially Private Distributed Computation via Public-Private Communication Networks
Lei Wang,Yang Liu,Ian Manchester,Guodong Shi
DOI: https://doi.org/10.48550/arXiv.2101.01376
2021-01-05
Distributed, Parallel, and Cluster Computing
Abstract:This paper studies the problem of multi-agent computation under the differential privacy requirement of the agents' local datasets against eavesdroppers having node-to-node communications. We first propose for the network equipped with public-private networks. The private network is sparse and not even necessarily connected, over which communications are encrypted and secure along with the intermediate node states; the public network is connected and may be dense, over which communications are allowed to be public. In this setting, we propose a multi-gossip PPSC mechanism over the private network, where at each step, randomly selected node pairs update their states in such a way that they are shuffled with random noise while maintaining summation consistency. We show that this mechanism can achieve any desired differential privacy level with any prescribed probability. Next, we embed this mechanism in distributed computing processes, and propose privacy-guarantee protocols for three basic computation tasks, where an adaptive mechanism adjusts the amount of noise injected in PPSC steps for privacy protection, and the number of regular computation steps for accuracy guarantee. For average consensus, we develop a PPSC-Gossip averaging consensus algorithm by utilizing the multi-gossip PPSC mechanism for privacy encryption before an averaging consensus algorithm over the public network for local computations. For network linear equations and distributed convex optimization, we develop two respective distributed computing protocols by following the PPSC-Gossip averaging consensus algorithm with an additional projection or gradient descent step within each step of computation. Given any privacy and accuracy requirements, it is shown that all three proposed protocols can compute their corresponding problems with the desired computation accuracy, while achieving the desired differential privacy.