Authors:

Adam Piaseczny

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

As the popularity of federated learning grows, new frameworks for decentralized settings are becoming widespread. By leveraging the advantages of environments without centralized coordinators they allow for fast and energy-efficient inter-client communication. However, these advancements also pose new security challenges, particularly in settings where malicious agents may interfere with the learning process. This paper investigates these security implications. We examine how different node centralities affect the potency of attacks in fully decentralized settings over various real-world and synthetic random directed graphs. Using established decentralized aggregation frameworks, we explore the relationship between the decrease in testing accuracy, optimality gap, convergence rate and attacks on the top-k most central nodes based on different centrality measures. We then look at how the adversarial distortion spread, and convergence rates are affected by various malicious agent eigenvector centralities. The results of this investigation will provide valuable insights into the spread of model alteration in distributed learning environments, a critical aspect of mitigating the impact of malicious interference in Federated Learning. The results also lay the groundwork for potential future work in the detection and identification of malicious interference, which would contribute to the development of more secure federated Learning frameworks.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Adam Piaseczny