Enhanced Wireless Interference Recognition via Federated Learning With Semi-Random Regularization Techniques

This paper presents an innovative approach to Wireless Interference Recognition (WIR) in communication systems, employing a deep learning (DL) framework. DL-based WIR methods have gained popularity for their precise classification abilities, yet they often face challenges with limited training samples. Addressing this, our research explores training a WIR network using multiple, distributed datasets that are constrained in size and challenging to share due to privacy concerns. We introduce a federated learning-based WIR method that trains multiple local networks and aggregates them for enhanced global optimization. This method contrasts with traditional centralized learning by only exchanging model weights between local clients and the server, significantly mitigating the risk of data leakage. To bolster the performance of local training with limited samples, we incorporate data augmentation and regularized training. Furthermore, we integrate a semi-random mechanism into the regularized training process, enabling a more comprehensive and effective feature learning from samples. Simulation results affirm that our proposed WIR method outperforms other advanced methods in recognition accuracy. Additionally, it confirms the semi-random mechanism's efficacy in improving training robustness and recognition accuracy, marking a significant advancement in DL-based WIR methodologies under constrained training conditions.