UFed-GAN: Secure Federated Learning over Wireless Sensor Networks with Unlabeled Data

The rising demand for deploying low-latency data analysis and protecting privacy in a cloud-based setting has led to the emergence of federated learning (FL) as an important learning paradigm over wireless sensor networks. Due to the success of FL, generative models such as generative adversarial networks (GANs) are now utilized in FL to provide higher privacy and utility. However, existing GAN-based FL approaches are power-hungry which poses unbearable demands on resource-limited distributed users. Considering practical learning systems involving limited computational power and unlabeled data over wireless networks, this work investigates FL in a resource-constrained and label-free data environment. Specifically, we propose a novel framework known as UFed-GAN that captures sensor-side data distribution without local classification training. We analyze the convergence and privacy of the proposed UFed-GAN. Our experimental results demonstrate the strong potential of UFed-GAN in addressing limited computational resources and unlabeled data while preserving privacy.