Robust Federated Learning with Valid Gradient Direction for Cloud-Edge-End Collaboration in Smart Grids
Bin Qian, Yuming Zhao, Jianlin Tang, Zhenshang Wang, Fusheng Li, Jing Wang, Shan He, Junlong Liu · 2024
With the advancement of carbon emissions reduction initiatives, there is a rapid increase in the current demand for electrical power, which necessitates a more efficient and reliable grid to meet this demand. Consequently, smart grids are proposed to address these new challenges. At the same time, to protect user privacy during big data analysis, smart grids introduce the federated learning algorithm involving three types of devices: cloud, edge, and end devices. However, previous federated learning algorithms based on cloud-edge-end collaboration typically do not assess whether user data contains noise, which leads to insufficient robustness of the models. To address this issue, we propose a new Robust Federated Learning (RoFed) method for end-device parameters selection based on the gradient directions of parameters. RoFed can filter out abnormal end devices based on the angle between gradient directions of each device and the average gradient direction per round. This filtering can prevent the end devices trained with noisy data from participating in the aggregation of parameters on the cloud device. Experiments on 13 real-world electric power data classification datasets demonstrate the effectiveness of RoFed.