Privacy-Preserving Distributed Processing Over Networks
Qiongxiu Li · 2021
Privacy has become a primary concern in modern world.Addressing the privacy issue is particularly challenging in the context of distributed processing due to many constraints such as absence of centralized coordination, limited computational resources and inevitable information exchange between different computing units.This thesis discusses on how to conduct signal processing over a network in a distributed manner without violating privacy.In particular, we focus on practical privacy-preserving solutions that are lightweight in terms of communication and computational cost.We first investigate existing privacy-preserving approaches which apply well-established cryptographic techniques into distributed processing tools.Secondly, we propose a new subspace perturbation based method that, instead of applying existing cryptographic techniques, directly exploits the potential of distributed processing tools such as distributed optimization for privacy-preservation.The proposed method is able to alleviate two fundamental limitations in existing approaches: the privacy-accuracy trade-off of differential privacy approaches and expensive communication cost incurred in secret sharing based approaches, respectively.Thirdly, based on the observation that all the above-discussed algorithms use the idea of inserting noise to mask the private data for privacy-preservation, we propose a new information-theoretical metric that is able to relate and compare all of them in a unified framework.Fourthly, we observe that there is typically a trade-off between the communication cost and privacy in noise insertion approaches and propose to address this trade-off, by making use of a quantization scheme in a particular way that the accuracy of the algorithm output is not deteriorated.Finally, continuing with the idea of exploring the potential of existing distributed processing tools for privacy-preservation, we take the first step to investigate the emerging graph signal processing tool and propose a privacy-preserving distributed graph filtering solution using noise insertion.This proposed solution has comparative performance compared with the above proposed subspace perturbation based distributed optimization approaches.