Data-Driven and Privacy-Preserving Cooperation in Decentralized Learning

Decentralized learning scenarios offer the opportu- nity of a flexible cooperation between learning nodes; in other words, each node may cooperate with an arbitrary subset of its peers. In such scenarios, we tackle the problem of choosing the nodes that cooperate towards the training of a machine learning model, hence, tweaking the cooperation graph connecting the nodes themselves. We propose and evaluate a data-driven approach to the problem, by proposing three metrics to choose the edges to activate in the cooperation graph, and an efficient iterative algorithm exploiting them. Through our performance evaluation, which leverages state-of-the-art datasets and neural network architectures, we find that privacy-preserving metrics accounting for the difference between local datasets are very effective in identifying the best edges to activate to improve the efficiency of model training without hurting performance.