New method improves privacy and accuracy in wireless networks
A new design uses wireless noise to protect data while training models across distributed access points.
Researchers developed a way to perform privacy-preserving federated learning over cell-free massive MIMO systems. This method uses over-the-air aggregation to improve performance.
Distributed access points (APs) work together to serve user equipments (UEs) through user-centric clustering. User-centric clustering involves APs collaborating to serve specific users. Over-the-air (OTA) aggregation happens when updates are transmitted over a shared uplink for averaging at a central processing unit.
The design uses clipped local model updates perturbed with Gaussian noise. Thermal noise from wireless reception can act as a privacy-masking resource. This can reduce or eliminate the need for artificial noise.
The researchers solved a problem to minimize aggregation variance. They used alternating optimization with difference-of-convex power allocation and greedy coordinate-descent clustering. This approach achieved up to 89% lower aggregation variance than conventional baselines.
The system enabled 97% test accuracy on MNIST and showed gains on CIFAR-10. Under strict privacy budgets, competing schemes remain near random-chance performance.
Key facts
- The method achieves up to 89% lower aggregation variance than conventional baselines.
- The design enables 97% test accuracy on MNIST.
- Thermal noise can act as a privacy-masking resource to reduce the need for artificial noise.