Doctoral Dissertations

Abstract

Federated Learning (FL), which facilitates collaborative model training and protects users' privacy, has drawn great interest from the research community. With FL, the participants train their models on local data and submit the corresponding updates for aggregation to a server. While concealing the participants' identities, FL may attract adversaries aiming to hamper the underlying model. These adversaries aim to submit malicious weight updates that corrupt the performance of the server model. Further, these models when communicated to the participating clients extend the behavior which is undesirable. Additionally, FL suffers from increased energy consumption at the edge device level due to local training and data collection along with communication overheads. In scenarios where the edge devices start with limited energy budget and train in heterogeneous settings, efficient client selection is needed for maximizing setup performance. Inefficient client behavior can lead to reduced setup lifetime and accuracy.

In this research, we aim to address the detection and isolation of such adversaries aiming to provide malicious weight updates in as targeted attacks on the setup without breaking the FL protocol and without compromising on the client data privacy. To perform this, we leverage model inspection on the submitted weights and use similarity between the client models. We then showcase that graph theoretic algorithms can be used to identify targeted attackers from the calculated pairwise similarity. We further discuss scenarios where using model similarity between the clients fails. Thereafter, we suggest alternative solutions for the scenarios described. Following this, we address the issue of energy-efficiency in edge devices in heterogeneous FL settings. We define efficient client behavior for data collection and selection followed by modeling local energy consumption and communication bandwidths. Additionally, we showcase how efficient client selection can increase the setup lifetime and performance with energy constraints.

Advisor(s)

Das, Sajal K.

Committee Member(s)

Tripathy, Ardhendu S.
Bhattacharjee, Shameek
Chatterjee, Shubham
Zhang, Yanzhi

Department(s)

Computer Science

Degree Name

Ph. D. in Computer Science

Publisher

Missouri University of Science and Technology

Publication Date

Spring 2026

Pagination

xii, 116 pages

Note about bibliography

Includes_bibliographical_references_(pages 107-115)

Rights

© 2026 Priyesh Ranjan , All Rights Reserved

Document Type

Dissertation - Open Access

File Type

text

Language

English

Thesis Number

T 12605

Download

Share

 
COinS