Improved Modulation Recognition Using Personalized Federated Learning

There are various types of signals around us and recognizing the signals allows a network system to effectively and efficiently transmit from the sender to the receiver. Signals can be appropriately decoded by modulation classification for data integrity. However, traditional digital network systems require devices to share their signal data to a remote server for modulation classification, resulting in high communication overhead and data privacy concerns. Federated learning can help to solve these problems, but it still has some limitations. Existing works mostly consider homogeneous FL settings for modulation recognition with independent and identically distributed (IID) data distributions and equal contributions of clients in jointly training the global modulation model. Different from the literature, this paper studies a heterogeneous modulation recognition network problem using a new personalized federated learning approach. The proposed scheme allows clients with different data sizes, data points, and computational abilities to participate in the training process. The algorithm is then used to do the local model training and evaluate their performance. An adaptive dynamic global model aggregation is developed where the local model's gradient is updated based on their performance with available learning rates. Simulation results demonstrate the superior performance of our approach over the state-of-the-art scheme, with an overall accuracy of 62.1% for non-IID data and 68.3% for IID data under varying signal-to-noise ratio settings.