FRAIM: A Feature Importance-Aware Incentive Mechanism for Vertical Federated Learning

Federated learning, a new distributed learning paradigm, has the advantage of sharing model information without revealing data privacy. However, considering the selfishness of organizations, they will not participate in federated learning without compensation. To address this problem, in this paper, we design a feature importance-aware vertical federated learning incentive mechanism. We first synthesize a small amount of data locally using the interpolation method at the organization and send it to the coordinator for evaluating the contribution of each feature to the learning task. Then, the coordinator calculates the importance value of each feature in the dataset for the current task using the Shapley value method according to the synthetic data. Next, we formulate the process of organization participation in the federation as a feature importance maximization problem based on reverse auction which is a knapsack auction problem. Finally, we design an approximate algorithm to solve the proposed optimization problem and the solution of the approximation algorithm is shown to be 1/2 -approximate to the optimal solution. Furthermore, we prove that the proposed mechanism is truthfulness, individual rationality, and computational efficiency. The superiority of our proposed mechanism is verified through experiments on real-world datasets.