LrFedIF: Low-Resource Federated Learning Based on Fingerprint Feature Imitation for Signal Recognition in Non-IID Scenarios

Federated learning is an emerging distributed machine learning method, and it is one of the most potential solutions to data privacy security issues. However, federated learning is currently facing a thorny challenge: in low-resource scenarios such as unbalanced data distribution, limited data quality, and missing labeling information, the performance of federated learning degrades severely. Especially when the data of each client is non-independent and identically distributed (Non-IID), the global aggregation model will suffer from severe model drift. In this paper, we propose a low-resource federated learning method based on fingerprint feature imitation (LrFedIF) for signal recognition tasks. A heterogeneous robust global classifier is trained by using the KL distance between feature prediction values to achieve the purpose of feature space alignment and effectively alleviate the problem of global model drift.