An Efficient Federated Automatic Modulation Recognition Scheme With Adaptive Model Pruning

The Internet access of edge sensors is greatly bolstered by the rapid evolution of information technology, thereby facilitating the proliferation of various communication devices and underscoring the need for efficient data transmission. Automatic modulation recognition (AMR) has played an increasingly significant role in radio frequency sensing, helping in signal and spectrum management, as well as efficient data transmission. To alleviate the heavy storage burdens, high computational costs, and potential disclosure of data privacy inherent in traditional centralized modulation methods, an innovative federated modulation recognition scheme designed for distributed scenarios is proposed in this article. It empowers multiple sensors to collaboratively train a global model without transmitting their raw datasets to a central server. Meanwhile, given that the local data from edge sensors often exhibit a nonindependent identically distribution, the adaptive model pruning is integrated to tackle the challenges arising from such data heterogeneity. Specifically, first, after receiving the broadcasted global model, each participating sensor employs adaptive model pruning in the training of a local model based on its specific heterogeneous data. Subsequently, the updated local model parameters are uploaded and then aggregated in the server. The above steps are reiterated until the global model achieves convergence. The experimental results demonstrate that the global model obtained through this proposed scheme can effectively identify modulation types within heterogeneous data, concurrently reducing both computational and communicative costs.