| Federated learning is a novel distributed machine learning technology,which allows mobile devices(e.g.,e.g.,smart phones and vehicular computing platforms)to hold raw training data on premises,train the target model locally,and only communicate the model update to a logically centralized server(e.g.,a cloud or edge data center)for aggregation in an iterative manner,which solves the shortage of training resources,lack of data and privacy disclosure.However,mobile device users’ participation in federated learning consumes computing cost,communication cost and time cost.At present,federated learning has no guarantee for the benefits of participants,thus they lack the motivation for federated learning.In addition,due to the resource gap between different users,which may deteriorate the performance of federated learning,so unreasonable user selection will also lead to unnecessary waste of time and resource cost.To tackle this issue,this work is committed to using randomized auction and online optimization algorithm design to incentives mobile devices to participate in federated learning and make reasonable user selection,so as to maximize the interests of participants.Specifically,this study regards the core server publishing federated learning tasks as the auctioneer and the mobile device as the bidder.Each round of federated learning task is used as an auction to collect bidding data(including bidding price,local data volume,local training accuracy and other relevant information)from all mobile devices within the communication range.Then,considering the heterogeneity of devices,user selection is carried out under the convergence of trained model and the longterm energy budget of mobile devices.The design of such an auction mechanism can more flexibly adapt to the dynamic market environment and meet the realtime needs of users when compared with the alternatives that allows mobile devices to price their data directly.For solving the user decision-making problem in this auction mechanism,this study constructs an online optimization problem of minimizing the longterm social cost in federated learning.The social cost refers to all costs incurred in the training process of all roles involved in the system,including mobile devices and the core server.The scenario we consider is that the server and multiple mobile devices are in a dynamic market,and the user’s data volume,training accuracy,communication energy consumption and computing energy consumption are confidential to other devices.Considering the limited resources and reducing the number of iterations as much as possible while ensuring the accuracy of the training model,this thesis takes the long-term energy consumption,model accuracy and sufficient amount of training data as constraints;uses the minimization of total cost(including communication and computing costs)generated by the server and all mobile devices in training process(global iteration)as the objective to construct the optimization problem,and then solve the value of decision variables that meet all constraints,that is,the mobile devices participating in the auction.Finally,this thesis constructs a federated learning simulation platform to complete the interaction between the server and multiple clients,including parameter transfer,model aggregation,parameter update and other functions,so as to truly reproduce the overall process of federated learning;At the same time,it completes the integration of optimization algorithm,auction mechanism and federated learning and training process,thus realize the whole mechanism studied in this thesis.Besides,by comparing the performance of federated learning training result and auction result with other alternatives,it shows the superiority of this proposed algorithm,and also verifies that the auction mechanism achieves the desired economic properties of truthfulness and individual rationality through the relevant experimental design. |
