垂直联邦学习中特征维度即服务的真实高效拍卖机制设计

Design of truthful and efficient auction mechanisms for feature-dimension-as-a-service in vertical federated learning

考虑垂直联邦学习(vertical federated learning,VFL)架构,将特征维度作为资源,进行了拍卖机制的设计。该机制包括一个可信任的第三方中央控制端作为拍卖方,以及若干任务需求者作为竞标方。为了防止竞标者进行欺诈性投标,首先提出了一种基于Vickrey-Clarke-Groves(VCG)机制的最优特征维度即服务(Optimal-FDaa S)机制,并通过整数线性规划确定拍卖的赢家。该机制能够最大化社会效益,但是计算复杂度较高。为了降低计算复杂度,提出了贪心特征维度即服务(Greedy-FDaaS)机制。以上两种机制都能保证3种性质,即真实性、个体理性与效率性。最后,通过仿真分析证明了所提机制的有效性。

The vertical federated learning(VFL)architecture was considered,and an auction mechanism using fea-ture dimensions as resources was designed.The mechanism included a trusted third-party central control as the auc-tioneer and several task demanders as bidders.To prevent fraudulent bidding,the optimal fea-ture-dimension-as-a-service(Optimal-FDaaS)based on the Vickrey-Clarke-Groves(VCG)mechanism was intro-duced,and the auction winner was determined through integer linear programming.This mechanism maximized so-cial benefits but had high computational complexity.To reduce this,greedy-feature-dimension-as-a-service(Greedy-FDaaS)was proposed.Both mechanisms ensured three properties:truthfulness,individual rationality,and efficiency.Finally,simulation analysis was provided to prove the effectiveness of the proposed mechanisms.