Session: 02-01: AI for Energy Sustainability I

Paper Number: 156364

156364 - An Enhanced Federated Learning Framework for Predictive Maintenance of Distributed Wind Systems 

Abstract: 

Background: As the demand for energy increases, distributed wind systems (DWSs) have arisen as a topic of interest over recent years. Numerous wind farms have arisen all over the world to contribute to this demand. As wind turbine technology advances, the effect of DWSs on the power grid becomes inevitable. Consequently, the emergence of DWSs has gained the attention of malicious groups and individuals. Cyber-attacks have become an increasing risk to DWSs. Due to underdeveloped cybersecurity measures, attackers can gain access to a single wind turbine and use the power and communication interconnections to gain access as far as power grids. Malicious attacks commonly target the supervisory control and data acquisition (SCADA) system to gain unwarranted control of wind farms to disrupt and destroy the power system. To help protect from attacks on the SCADA system, federated learning (FL) frameworks with trusted execution environment (TEE) implementation pose a solution. However, current TEEs-based FL frameworks face several significant challenges that hinder their application and effectiveness. Those FL frameworks face one challenge that the server, which is responsible for model aggregation and distribution, is unable to defend itself from malicious physical attacks and memory-access pattern (MAP) attacks. Secondly, the capability of FL clients to install and maintain TEEs is limited due to factors like hardware limitations, device heterogeneity, and computational resources. This study addresses these issues by proposing a server-focused FL framework to mitigate the need for extensive client-side TEE attestations. Additionally, we aim to enhance the server’s resilience against side-channel attacks.

Method: Three solutions are developed to enhance the use of TEEs in an FL framework for DWSs. 1) Exclusive use of TEEs on the server level. By concentrating TEEs on the server, the extensive verification of potential untrusted clients is mitigated. Therefore, operational complexity is reduced while the overall scalability of FL frameworks for DWSs is enhanced. 2) Implement a TEE-supported mutual attestation. Before proceeding with model updates, both the server and clients engage in a cryptographic protocol to verify each other’s authenticity and integrity before model gradients and other data is exchanged. This diminishes the submission of compromised updates. In addition, a Byzantine fault tolerance (BFT) protocol is implemented to ensure that an FL framework can still converge in the presence of malicious behavior under a certain threshold. 3) Integrating Oblivious RAM (ORAM). MAP attacks observe predictable memory access patterns to infer and reverse engineer sensitive data. To combat this, ORAM can obfuscate the server’s memory access patterns to secure processed data from side-channel vulnerabilities.

Conclusion: These solutions aspire to strengthen data privacy, integrity, and the reliability of FL frameworks. The proposed approach provides a robust solution without compromising client-side performance, creating a precise balance between practicality and security. The proposed enhanced FL framework has been validated through considerable amounts of experimentation that measure key metrics like latency, system security, and reliability of the FL system. These metrics confirm that our approach provides a balanced trade-between practicality and high security making it effective for real-world applications.

Presenting Author: William Mayfield Mississippi State University

Presenting Author Biography: William Mayfield is a graduate student at Michael W. Hall School of Mechanical Engineering at Mississippi State University. His research area is AI-based predictive maintenance for distributed wind systems.