Decentralized Online Bandit Federated Learning Over Unbalanced Directed Networks

This paper investigates a class of privacy-enhancing decentralized federated learning (DFL) algorithms. The majority of current DFL algorithms rely on the premise that obtaining gradient information for the cost function is efficient and that the weight matrix of the communication network adheres to either doubly stochastic or column stochastic. However, in many cases, obtaining gradient information may be challenging, and the requirement for the weight matrix is also stringent. To overcome these challenges, a decentralized online bandit federated learning algorithm with differential privacy (DP) is proposed. By employing the one-point bandit feedback (OPBF), the algorithm can estimate the gradient without the requirement of complex computation of gradient information. Moreover, the proposed algorithm models the communication network as an unbalanced directed network with a row stochastic weight matrix, eliminating the stringent requirements on the network. Our algorithm accomplishes both sublinear convergence and
$\varepsilon$
-DP through rigorous theoretical derivations. Finally, the effectiveness of our algorithm is validated through simulation experiments on multiple mainstream datasets with various data scenarios.