基于人在回路的纵向联邦学习模型可解释性研究OACSTPCD

Vertical federated learning(VFL)is commonly used for cross-domain data sharing in high-risk scenarios.Us-ers need to understand and trust model decisions to promote the application of models.Existing research primarily fo-cuses on the trade-off between explainability and privacy within VFL,and fails to fully meet the needs of users for estab-lishing trust and fine-tuning models.To address these issues,we proposed an explainable vertical federated learning method based on human-in-the-loop(XVFL-HITL),which incorporated user feedback into the VFL's Shapley value-based explainability approach through a distributed HITL structure,using the knowledge of all VFL participants to correct training data and enhance model performance.Furthermore,considering privacy concerns,this paper employed the addi-tive principle of Shapley values to integrate the feature contribution values of all entities other than the target participant into an aggregated measure,which effectively protected the feature privacy of each participant.Experimental results indi-cated that on benchmark data,the explainability results of XVFL-HITL were effective and could well protect the feature privacy of user.Additionally,compared to VFL-Random and VFL-Shapley,the model accuracy of XVFL-HITL improved by approximately 14%and 11%,respectively.