| PhD Viva

Name of the Speaker: Mr. Shamik Bhattacharyya (EE18D005)
Guide: Dr. Rachel Kalpana K
Venue: Online
Online meeting link: https://meet.google.com/rji-pzvq-ota
Date/Time: 8th July(Monday), 2024, at 11.15AM
Title: Resilient Algorithms for Distributed Estimation

Abstract :

The growth in the use of large-scale systems has led to an increasing interest in developing more distributed approaches. A widely used framework to design such distributed approaches is a multi-agent system (MAS) comprising autonomous entities, called agents, equipped with computation and communication functionality. In this work, we focus on designing consensus-based algorithms for distributed estimation of parameters in MASs. The distributed nature of operation in large-scale MAS with agents spread out geographically poses a challenge in ensuring the proper functioning of every agent. An area of increasing concern is attacks on vulnerable agents by external entities, called adversarial attacks, which intend to prevent the MAS from fulfilling its global objective by compromising a part or the entire functionality of those agents. Thus, alongside developing new distributed algorithms for MAS, it is also necessary to design resilience for existing distributed approaches to such adversarial attacks. We first address the distributed estimation of a static parameter in a MAS where the agents are equipped with sensors that collect measurements of the required parameter. We consider that some of the agents’ sensors are under false data injection type attack, which manipulates the sensed data, thereby trying to cause the MAS to estimate the parameter incorrectly. Moreover, we consider the communication among the agents to be unidirectional, represented by directed graphs, which provides an additional challenge for designing consensus-based algorithms. We propose a novel REWB (Resilient Estimation through Weight Balancing) algorithm that combines the consensus+innovations approach to ensure resilience to the sensor attacks, and the concept of weight balancing to address the issue of unidirectional communication. Next, we focus on the distributed estimation of a dynamic parameter in a MAS, where the agents collect measurements of locally available dynamic reference signals and collaboratively track the average of these time-varying signals. For this problem, we consider a broader scope of adversarial attack, namely Byzantine attack, which compromises the entire functionality of the agents. To address resilience to Byzantine adversaries, we invest in the security of a few agents, namely trusted agents, such that they are never compromised by adversaries. We propose a novel ResDAC (Resilient Dynamic Average Consensus) algorithm that utilises the non-compromising nature of the trusted agents to ensure that all the non-adversarial agents track the desired dynamic parameter within some small tolerance. Finally, we focus on federated learning, a decentralised approach to machine learning that uses a central server to aggregate local model estimates from the clients to generate a global model estimate. However, the presence of a single central server brings in the vulnerability of being a central point of failure. To provide resilience to such failures, we propose a distributed approach to federated learning consisting of multiple servers with corresponding disjoint sets of clients. The servers are capable of communicating among themselves, alongside the usual periodic communication with their corresponding clients. Further, we propose a novel DFL (Distributed Federated Learning) algorithm, designed on the concepts of distributed gradient descent for estimating the local client models, and consensus for agreement among the servers over a common global model. For each of the proposed novel algorithms, we develop convergence guarantees, and illustrate their performance through numerical simulations.