Smart grid security through fusion-enhanced federated learning against adversarial attacks

Abstract

The proliferation of smart grids introduces significant challenges for energy networks in managing and securing the vast data they generate. Federated Learning (FL) provides a cost-effective and privacy-preserving framework for distributed model training, addressing critical concerns around customer data privacy and security. However, FL remains vulnerable to adversarial threats, particularly data poisoning attacks, which can significantly impair model performance. This study presents a novel data poisoning attack and proposes a mitigation framework tailored for resource-constrained smart grids. The Centroid-Based Anomaly Aware Federated Averaging (CBAA-FedAvg) framework is introduced, achieving a Mean Absolute Percentage Error (MAPE) of 2.7 percent, closely aligning with baseline performance while maintaining robustness. CBAA-FedAvg integrates advanced fusion techniques, including parameter quantization (from 32-bit floating point to 8-bit fixed point) and dynamic clustering, to minimize computational complexity and optimize data processing. Furthermore, an automatic stopping criterion optimizes convergence, reducing energy consumption and computation time. Experimental results demonstrate that CBAA-FedAvg exhibits remarkable resilience to both data and model poisoning attacks, leveraging fusion strategies to enhance security and efficiency. This framework provides a scalable and effective solution for improving the fusion, security, and operational efficiency of smart grids.