Federated Multicriteria Learning: A Goal Programming Perspective

Federated Learning is a framework addressing the problem of decentralized learning when the data is scattered among numerous computing nodes. However, by sharing the gradient, Federated Learning approaches fail to comply with recent privacy protection regulations. In this paper, we propose a Goal Programming approach to address this issue. In particular, we define the federated model as a parallel ensemble in which the weights belong to the solution of a Weighted Goal Programming problem. The proposed approach is privacy compliant, as the nodes receive only the final prediction functions. The proposed approach also has the advantage of being model and metric agnostic. To test the performance of our approach, we employ a novel dataset, specifically for Federated Learning, the "Federated Extended MNIST". From the numerical experiment, it emerges that the federated Convolutional Neural Network performs much better than the Residual Network given the same training periods. Last, the Convolutional Neural Network architecture generalizes more on the edge nodes as the training increase slightly.