Skip to Main content Skip to Navigation

Energy and privacy aware estimation over adaptive networks

Abstract : Distributed estimation over adaptive networks takes advantage of the interconnections between agents to perform parameter estimation from streaming data. Compared to their centralized counterparts, distributed strategies are resilient to links and agents failures, and are scalable. However, such advantages do not come without a cost. Distributed strategies require reliable communication between neighbouring agents, which is a substantial burden especially for agents with a limited energy budget. In addition to this high communication load, as for any distributed algorithm, there may be some privacy concerns particularly for applications involving sensitive data. The aim of this dissertation is to address these two challenges. To reduce the communication load and consequently the energy consumption, we propose two strategies. The first one involves compression while the second one aims at limiting the communication cost by sparsifying the network. For the first approach, we propose a compressed version of the diffusion LMS where only some random entries of the shared vectors are transmitted. We theoretically analyse the algorithm behaviour in the mean and mean square sense. We also perform numerical simulations that confirm the theoretical model accuracy. As energy consumption is the main focus, we carry out simulations with a realistic scenario where agents turn on and off to save energy. The proposed algorithm outperforms its state of the art counterparts. The second approach takes advantage of the multitask setting to reduce the communication cost. In a multitask setting it is beneficial to only communicate with agents estimating similar quantities. To do so, we consider a network with two types of agents: cluster agents estimating the network structure, and regular agents tasked with estimating their respective objective vectors. We theoretically analyse the algorithm behaviour under two scenarios: one where all agents are properly clustered, and a second one where some agents are asigned to wrong clusters. We perform an extensive numerical analysis to confirm the fitness of the theoretical models and to study the effect of the algorithm parameters on its convergence. To address the privacy concerns, we take inspiration from differentially private Algorithms to propose a privacy aware version of diffusion LMS. As diffusion strategies relies heavily on communication between agents, the data are in constant jeopardy. To avoid such risk and benefit from the information exchange, we propose to use Wishart matrices to corrupt the transmitted data. Doing so, we prevent data reconstruction by adversary neighbours as well as external threats. We theoretically and numerically analyse the algorithm behaviour. We also study the effect of the rank of the Wishart matrices on the convergence speed and privacy preservation.
Complete list of metadatas

Cited literature [100 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, April 2, 2020 - 11:35:38 AM
Last modification on : Monday, October 12, 2020 - 11:10:20 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02529305, version 1



Ibrahim El Khalil Harrane. Energy and privacy aware estimation over adaptive networks. Distributed, Parallel, and Cluster Computing [cs.DC]. COMUE Université Côte d'Azur (2015 - 2019), 2019. English. ⟨NNT : 2019AZUR4041⟩. ⟨tel-02529305⟩



Record views


Files downloads