Federated learning with self-updating server parameters for smart home intrusion detection in Non-IID environments

Federated Learning (FL) has been widely applied in Intrusion Detection Systems (IDS) for the smart home due to its advantages in data privacy protection. However, as the variety and number of smart home devices increase, the phenomenon of Non-Independent and Identically Distributed (Non-IID) data among clients becomes more pronounced. In this situation, traditional FL, which relies on model parameters aggregation, may lead to client drift and subsequently degrade model performance. Therefore, this paper proposes a Smart Home Intrusion Detection Model Based on Federated Learning with Self-Updating Server Parameters (FL-SUSP). Firstly, FL-SUSP generates soft labels for the unlabeled distillation dataset and utilizes the domain discriminator to determine the weights of the soft labels so that the client can effectively distinguish data of different familiarity levels. Secondly, FL-SUSP designs a self-labeling method based on weighting soft voting on soft labels, which improves the quality of labels by integrating different clients' understanding of the data to generate more accurate final labels for the distillation dataset. Finally, FL-SUSP selects superior local models based on performance advantages for aggregation and initializes the server model with the aggregated results, followed by parameters-self-updating according to the labeled distillation datasets after self-labeling. In this way, the impact of Non-IID data on the model is alleviated, and model performance is improved. This paper conducted experiments in three Non-IID scenarios. The experimental results indicate that under optimal conditions, FL-SUSP has a maximum improvement of 30.88% compared to the baseline model. In addition, FL-SUSP achieves stable intrusion detection performance on clients while maintaining high accuracy.