[论文解读] A Federated Learning Framework for Smart Grids: Securing Power Traces in Collaborative Learning

Haizhou Liu, Xuan Zhang+2 more|arXiv (Cornell University)|Mar 22, 2021

Privacy-Preserving Technologies in Data参考文献 21被引用 42

一句话总结

本文提出面向智能电网的水平联邦学习和垂直联邦学习框架，以在不暴露个人痕迹的前提下安全学习用电模式，采用 Paillier 加密；通过在 Zhuhai 的案例研究，使用 LSTM、VFLR 和 SecureBoost 验证无损且保护隐私的学习。

ABSTRACT

With the deployment of smart sensors and advancements in communication technologies, big data analytics have become vastly popular in the smart grid domain, informing stakeholders of the best power utilization strategy. However, these power-related data are stored and owned by different parties. For example, power consumption data are stored in numerous transformer stations across cities; mobility data of the population, which are important indicators of power consumption, are held by mobile companies. Direct data sharing might compromise party benefits, individual privacy and even national security. Inspired by the federated learning scheme from Google AI, we propose a federated learning framework for smart grids, which enables collaborative learning of power consumption patterns without leaking individual power traces. Horizontal federated learning is employed when data are scattered in the sample space; vertical federated learning, on the other hand, is designed for the case with data scattered in the feature space. Case studies show that, with proper encryption schemes such as Paillier encryption, the machine learning models constructed from the proposed framework are lossless, privacy-preserving and effective. Finally, the promising future of federated learning in other facets of the smart grid is discussed, including electric vehicles, distributed generation/consumption and integrated energy systems.

研究动机与目标

• Address privacy and data ownership concerns in smart grid analytics across distributed parties.
• Develop horizontal federated learning (HFL) for collaborative power forecasting across locations.
• Develop vertical federated learning (VFL) sub-frameworks for cross-party feature-based predictions.
• Demonstrate lossless and privacy-preserving learning under encrypted parameter exchanges.

提出的方法

• Use horizontal federated learning with Paillier-encrypted model updates to train a shared model without exchanging raw data.
• Apply LSTM as the unanimous model for horizontal power consumption forecasting across transformer stations.
• Implement vertically-federated linear regression (VFLR) and SecureBoost for vertically distributed data to predict power consumption.
• In VFLR, employ a three-party setup with masking and encryption to securely update parameters.
• In SecureBoost, leverage encrypted gradient statistics and a third-party decryption pattern to perform privacy-preserving split finding.
• Evaluate convergence and privacy preservation using mean squared error (MSE) as the evaluation metric.

实验结果

研究问题

• RQ1 Can horizontal and vertical data separation in smart grids be effectively learned through federated frameworks without leaking raw traces?
• RQ2 Do HFL and VFL approaches achieve lossless performance comparable to centralized training under encryption?
• RQ3 How do encryption-based federated methods perform under practical communication delays and data withdrawal scenarios?
• RQ4 What is the impact of using nonlinear models (SecureBoost) versus linear models (VFLR) in vertically partitioned data?

主要发现

• Horizontal FL converges within 10 epochs with MSE around 0.04.
• With 40% communication delay, HFL still converges within 10 epochs but final MSE rises to about 0.051.
• Vertical FL with linear regression (VFLR) converges from an initial MSE of 0.92 to about 0.28; withdrawal of mobility data raises MSE to about 0.38.
• SecureBoost (vertically-federated XGBoost) achieves lower MSE, converging to about 0.14 with more trees, outperforming VFLR, and exhibits higher robustness to early data withdrawal with MSE around 0.28.
• VFL frameworks are lossless, matching centralized performance in the reported setups; SecureBoost maintains privacy while achieving competitive predictive accuracy.