Indoor intrusion detection is essential for various applications, including security systems and smart homes. Recently, WiFi-based detection has gained popularity due to its low cost and non-invasive nature. Current Channel State Information (CSI) based frameworks primarily use deep learning to extract gait signatures; however, their performance depends heavily on extensive labeled datasets. These methods struggle to differentiate between unlabeled and labeled data that exhibit similar features. To address this challenge, we propose a novel Two-level Feature Fusion model for Indoor Intrusion Detection (TFF-IID) utilizing commercial WiFi CSI. The model adopts a two-level structure to learn rich feature representations and introduces a Transformer with multi-head self-attention alongside a multi-scale convolution module to process sensor data. Additionally, it incorporates a self-supervised learning module to capture general normality patterns. Based on this architecture, TFF-IID achieves accurate intrusion detection using only CSI. Empirical evaluations on a private gait dataset demonstrate that TFF-IID achieves an intrusion detection accuracy of 73.5% and an F1-score of 76.2% across 10 unauthorized subjects. Moreover, cross-scenario assessments verify that the proposed model maintains high efficiency and robustness in environments characterized by diverse spatial layouts and multipath complexities. Furthermore, TFF-IID outperforms the best baseline by 19.7% and 25.7% in accuracy and F1-score, respectively.

The influence of non-Independent Identically Distribution (non-IID) data on Federated Learning (FL) has been a serious concern. Clustered Federated Learning (CFL) is an emerging approach for reducing the impact of non-IID data, which employs the client similarity calculated by relevant metrics for clustering. Unfortunately, the existing CFL methods only pursue a single accuracy improvement, but ignore the convergence rate. Additionlly, the designed client selection strategy will affect the clustering results. Finally, traditional semi-supervised learning changes the distribution of data on clients, resulting in higher local costs and undesirable performance. In this paper, we propose a novel CFL method named ASCFL, which selects clients to participate in training and can dynamically adjust the balance between accuracy and convergence speed with datasets consisting of labeled and unlabeled data. To deal with unlabeled data, the prediction labels strategy predicts labels by encoders. The client selection strategy is to improve accuracy and reduce overhead by selecting clients with higher losses participating in the current round. What is more, the similarity-based clustering strategy uses a new indicator to measure the similarity between clients. Experimental results show that ASCFL has certain advantages in model accuracy and convergence speed over the three state-of-the-art methods with two popular datasets.