Adaptive Weighting via Federated Evaluation Mechanism for Domain Adaptation with Edge Devices

Abstract

Federated Learning is an emerging application paradigm of edge computing in smart cities. On the one hand, it enables efficient, private, and secure processing of sensitive data. On the other hand, it alleviates the burden of centralized data processing for the smart city. However, in real-world scenarios, performance degradation caused by domain adaptation has become a bottleneck that limits the widespread application of federated learning. Most existing approaches tackle the issue by designing novel local learning approaches to transfer knowledge among different domains while ignoring the optimization for global model aggregation. To address this issue, we propose a novel approach that leverages the label-free adversarial learning technique to evaluate the representations learned by the different domains under the global model. With the constraints of the federated setting, we minimize the discrepancy by aligning each distribution to the global distribution. Additionally, we have developed a fast detector to enhance the quality of the generated images. Through extensive experiments on image classification tasks, we have demonstrated promising results and shown that our approach can serve as a robust plugin for other local optimizers in Federated Learning.