A new federated learning framework lets wireless devices learn to recognize radio signal types together, without ever pooling their raw data.
The system, called FedSSL-AMC, tackles automatic modulation classification, the task of identifying what kind of radio signal a receiver is looking at, which underpins spectrum-aware wireless networks. Instead of sending raw I/Q signal recordings to a central server, edge devices jointly train a shared neural network encoder using self-supervision on unlabeled signals, then add a lightweight SVM classifier trained on just a handful of labeled examples per device. The researchers also proved convergence guarantees for their training procedure and derived a mathematical bound on how well the resulting classifier separates signal classes under noisy conditions. On synthetic and real over-the-air datasets, the approach outperformed standard supervised federated learning baselines in nearly every tested scenario, including uneven signal strengths, frequency drift, and skewed label distributions across devices.
Centralizing raw radio data is costly and a privacy liability, and models trained that way often break down once real-world channel conditions diverge from the training set. This method keeps data local and needs very few labels, which matters because labeling radio signals usually requires specialized equipment or expert time that most edge deployments do not have.
The gains are consistent rather than dramatic, and the SVM layer means the system still depends on someone providing a few labeled examples per class - fully label-free deployment remains a problem for a future paper.