考虑隐私保护的在线单点反馈无投影去中心化联邦学习算法OA

This study investigates a decentralized federated learning(DFL)algorithm that considers client privacy protection,with the objective of safeguarding private information of each client from expo-sure while ensuring the model converges to the global optimal solution.First,a Frank-Wolfe projection-free decentralized federated learning algorithm based on differential privacy is proposed for the iterative process.The algorithm is integrated with online single-point feedback technology to avoid complex projec-tion calculations under high-dimensional constraint sets,and solves the problem of inaccessible gradient information by approximating gradients with function values.In scenarios without a centralized server,the algorithm achieves client privacy protection.Theoretical analysis demonstrates that the algorithm can con-verge to the global optimal solution.Finally,the effectiveness of the algorithm is verified through simula-tion experiments on a dataset.