An Uncertainty-Aware Multi-Dimensional Auction Mechanism Based on Age of Update for Federated Learning

Federated learning (FL) can effectively address the phenomenon of data silos. It allows multiple clients to jointly train machine learning models and protect their local data confidentiality. However, it is a key challenge to stimulate the enthusiasm of these clients, ensure their willingness to invest in model training, and ensure that the model can efficiently reach a convergence state and maintain a high level of testing accuracy. This article proposes an uncertainty-aware auction mechanism (UAMTRA) that considers the training time, client reputation, and Age of Update (AoU). Firstly, client reputation is introduced to mirror truthfully the client's data quality and training status. AoU is introduced to measure the interval required for the server to receive the latest updates from the client to explore better and utilize the client's data resources and avoid overfitting. Subsequently, we reformulate the objective of maximizing social welfare as an NP-hard problem. To resolve this problem, we introduce the UAMTRA mechanism, which is capable of approaching maximal social welfare while ensuring minimal time complexity. In addition, UAMTRA also considers the differences in client network environments and heterogeneity of data resources, which can fully exploit and utilize client data resources to promote rapid convergence of FL and achieve higher testing accuracy. Ultimately, the efficacy of UAMTRA was confirmed through simulated experiments. Compared to other mechanisms, UAMTRA exhibited quicker convergence rates and superior testing accuracy on the MNIST datasets.