Double Auction Mechanism Design in Federated Learning

Zishang Chen, Juan Li

Abstract: In FL, participants cooperatively train a global model with their local data. The participants, however, may be heterogeneous in terms of data distribution. In such cases, FL might produce a biased global model that is not optimal for each participant. Moreover, participants may be reluctant to share their models due to privacy concerns and the incurred computation and communication costs. In this paper, we design a double auction mechanism to 1. recommend proper models for each participant’s model aggregation to ease distribution heterogeneity and improve his/her personal model performance (PMP) on the local data, 2. to make most model buyer-seller pair succeed and ensure truthfulness.