Rethinking Federated Learning: An Adversarial Perspective on Global vs. Local Learning for Load Forecasting

Manzoor, Habib Ullah, Shabiar, Attia, Nguyen, Dinh C, Mohjazi, Lina, Kaushik, Aryan ORCID: https://orcid.org/0000-0001-6252-4641 and Zoha, Ahmed (2024) Rethinking Federated Learning: An Adversarial Perspective on Global vs. Local Learning for Load Forecasting. In: 2024 IEEE Conference on Standards for Communications and Networking (CSCN), 25 November 2024 - 27 November 2024, Belgrade, Serbia.

Accepted Version Available under License In Copyright. Download (927kB)

Abstract

Resilient federated learning (FL) systems are essential for accurate load forecasting, especially when under adversarial attacks. Since these systems aggregate decentralized data from various sources, they are particularly vulnerable to attacks that can undermine forecast accuracy and reliability. To enhance robustness in load forecasting, our study investigates methods for strengthening FL systems by optimizing the balance between global and local learning processes. This paper explores the trade-offs between global and local learning in federated load forecasting under adversarial conditions. We develop a neural network framework tailored for federated short-term load forecasting and assess its performance against model poisoning attacks. Our experiments demonstrate that increasing the number of local training epochs while reducing global communication rounds can significantly enhance model robustness. Specifically, when local epochs are increased from 1 to 10 and global epochs are decreased from 1000 to 100, the average client Mean Absolute Percentage Error (MAPE) decreases from 92.3 % to 4.3 % under attack conditions. This improvement stems from a reduced attack surface and the concept of catastrophic forgetting, where local models gradually mitigate adversarial effects through extended training on authentic data, providing valuable insights for the design of secure and efficient distributed energy forecasting systems.