DASFAA 2024, LNCS 14850, pp. 409–424, 2024. VFDV-IM - An Efficient and Securely Vertical Federated Data Valuation._OceanBase.pdf

迹部景吾

16页

0次

2024-12-31

免费下载

VFDV-IM:AnEﬃcient and Securely

Vertical Federated Data Valuation

Xiaokai Zhou

,XiaoYan

,XinyanLi

,HaoHuang

,QuanqingXu

Qinbo Zhang

,YenJerome

,ZhaohuiCai

)

, and Jiawei Jiang

)

Wuhan University, Wuhan, Hubei Province, China

{xiaokaizhou,xinyan

li,haohuang,qinbo zhang,zhcai,jiawei.jiang}@whu.edu.cn

Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong SAR, China

OceanBase, Ant Group, China

xuquanqing.xqq@oceanbase.com

The University of Macau, Macau, China

jeromeyen@um.edu.mo

Abstract. Vertical federated learning enables multiple participants to

build a joint machine learning model upon distributed features of over-

lapping samples. The performance of VFL models heavily depends on the

quality of participants’ local data. It’s essential to measure the contribu-

tions of the participants for various purposes, e.g., participant selection

and reward allocation. The Shapley value is widely adopted by previ-

ous works for contribution assessment. However, computing the Shapley

value in VFL requires repetitive model training from scratch, incurring

exp ensive computation and communication overheads. Inspired by this

challenge, in this paper, we ask: can we eﬃciently and securely perform

data valuation for participants via the Shapley value in VFL?

We call this problem Vertical Federated Data Valuation, and intro-

duce VFDV-IM, a method utilizing an Inheritance Mechanism to expe-

dite Shapley value calculations by leveraging historical training records.

We ﬁrst propose a simple, yet eﬀective, strategy that directly inher-

its the model trained over the entire consortium. To further optimize

VFDV-IM, we propose a model ensemble approach that measures the

similarity of evaluated consortiums, based on which we reweight the his-

torical models. We conduct extensive experiments on various datasets

and show that our VFDV-IM can eﬃciently calculate the Shapley value

while maintaining accuracy.

Keywords: Vertical federated learning

· Data valuation

1Introduction

Building high-quality machine learning (ML) models usually requires collecting

and merging data from various organizations. However, due to data protection

X. Zhou and X. Yan—Equal contribution.

! The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024

M. Onizuka et al. (Eds.): DASFAA 2024, LNCS 14850, pp. 409–424, 2024.

https://doi.org/10.1007/978-981-97-5552-3

_28

BBAAD9C20180234D78A0072836F0B3B092B9B20912680BA0A4D98434B15B2B741B4BB438915BAB0F22C920089846A2EBF2E9218AC1D05B311BBFC26F7C7E3FD3241123ADB72279F764FF283767F74BC05967EB8175F7041048FA519C2ADBBCC8D9E620954E3

410 X. Zhou et al.

Fig. 1. Illustration of VFL tasks setting. In VFL, each participant possesses a distinct

set of features on the same set of samples, and only one participant holds the labels.

Through secure transmission and collaboratively computation, these participants can

train a global ML model over the joint feature space.

laws and increased privacy concerns, participants frequently limit data shar-

ing beyond their organization’s boundaries. As a distributed machine learning

scheme, federated learning (FL) can enable collaborative ML model training in

afederated“data consortium” in a secure and privacy-preserving way without

the need for centralized data aggregation. Based on data distribution, FL can be

categorized into horizontal federated learning (HFL) and vertical federated learn-

ing (VFL). In HFL, multiple participants contribute data samples that share the

same feature space. In contrast, VFL focuses on situations where diﬀerent par-

ticipants hold diﬀerent features of the same set of data samples.

Ver tical Fed erate d Learni ng . In this paper, we focus on VFL tasks. As

shown in Fig. 1, we consider a federated data consortium

P comprising m par-

ticipants, denoted by

P = {P

,...,P

}.Intheconsortium,participantsshare

identical data samples but possess unique feature spaces. For example, for the

same users within a city, an e-commerce platform collects shopping information

from customers and a bank that has personal ﬁnance data.

Data Valuation in FL. Due to diﬀerent data volumes and data qualities,

the contribution of each participant to the performance of the global model

varies in FL. Some participants who provide large-scale and high-quality data

may introduce signiﬁcant contributions to the trained model, while others with

redundant data or erroneous data may impede the global model from achiev-

ing satisfactory performance or even deteriorate the model. Thus, it’s critical

to evaluate the data quality of each participant, without sacriﬁcing the privacy

of participants’ local data. Many eﬀorts have been devoted to data valuation

for HFL [12,29]. For example, Fed-inﬂuence [24]quantiﬁestheinﬂuenceofeach

client in terms of model parameters and Zhang et al. [27] measure the contribu-

tion of participants’ data by computing the cosine similarity between local and

of 16

免费下载

paper oceanbase

关注

评论