THESIS
2023
1 online resource (xviii, 105 pages) : color illustrations
Abstract
The rapid advancement of networking technologies in distributed multi-agent
systems has raised concerns about data privacy, as messages containing sensitive
and private information transmitted between agents are vulnerable to
interception by adversaries. Meanwhile, due to limited communication bandwidth
and capacity, how to transmit less information without sacrificing system
performance has also gained considerable attention. In this thesis, we focus on
privacy-preserving and communication-efficient distributed optimization problems
in multi-agent systems. To be specific, we propose privacy-preserving algorithms
for distributed average consensus and cooperative optimization problems;
and introduce an information compressed scheme for distributed competitive games.
First, with the aware...[
Read more ]
The rapid advancement of networking technologies in distributed multi-agent
systems has raised concerns about data privacy, as messages containing sensitive
and private information transmitted between agents are vulnerable to
interception by adversaries. Meanwhile, due to limited communication bandwidth
and capacity, how to transmit less information without sacrificing system
performance has also gained considerable attention. In this thesis, we focus on
privacy-preserving and communication-efficient distributed optimization problems
in multi-agent systems. To be specific, we propose privacy-preserving algorithms
for distributed average consensus and cooperative optimization problems;
and introduce an information compressed scheme for distributed competitive games.
First, with the awareness of preserving agents’ privacy in a distributed system,
we develop a novel push-sum approach for directed networks that can protect
the privacy of all agents while achieving average consensus simultaneously. The paradigm utilizes a state decomposition mechanism to preserve the agents’
privacy. Second, for general objective functions, we propose a variant of the state
decomposition mechanism, where Laplacian noise is perturbed on the substate
of the gradient to achieve ∈-differential privacy for each agent. Furthermore,
linear convergence is guaranteed under a constant stepsize policy. Finally, for
communication-efficient distributed competitive games, we introduce a discrete-time
distributed NE seeking algorithm with information compressed by general
compressors before transmission. The proposed difference compression method
ensures linear convergence to the Nash equilibrium of the game. Moreover, the
communication cost is saved in terms of transmitting bits by adopting a general
class of compressors.
Post a Comment