Secure Coalition Formation for Federated Machine Learning

Federated machine learning allows multiple collaborative parties to build a machine learning model in a distributed fashion where the data is distributed among the parties. There can be few restrictions on the participation of such parties in a federate machine learning process. For example, parties may have spatial restrictions such that only a limited number of parties from one locality can participate in one federated machine learning process. Only a fixed number of parties may participate in one federated machine-learning process. To include such restrictions on federated machine learning we need to consider optimally of such group formation. This optimization condition will ensure that each group’s aggregated DNN model reaches sufficient accuracy. In this paper, we investigate such constraints on the federated machine learning process. Specifically, we explore party-affiliation game-based partition over the parties for a federated machine learning model. Partition function games find optimal and stable partitions (where no party is better off by changing groups to improve the accuracy of the aggregated DNN model) among the parties and one federated machine learning model is formed in each partition. Further, we investigate methods for secure and privacy-preserving federated machine learning processes where parties are partitioned into a set of groups and the machine learning model of one party is not shared with other parties in different groups. Our main contributions are (a) we proposed a federated machine learning protocol for a group of participants where the number of parties in each group may not exceed a predefined threshold, (b)we proposed a federated machine learning protocol that forms a stable partition among the parties such that no party can leave a group and join another group to improve the accuracy of aggregated DNN models, (c) Our proposed federated machine learning protocol is secure as it prevents a party from reporting the wrong DNN model to the aggregator, (d)our proposed federated machine learning protocol is secure as ensures that the aggregator correctly generates the coalition formation protocol, (e)our proposed federated machine learning protocol is secure and privacy-preserving as it ensures that valid DNN model parameters are exchanged among the parties.