HA-DP-FedAvg: Privacy Scheduling and Heterogeneity-Aware Aggregation for Federated Credit Risk Prediction

Zihang Zeng; Yunkun Shi; Kai Liu

HA-DP-FedAvg: Privacy Scheduling and Heterogeneity-Aware Aggregation for Federated Credit Risk Prediction

Zihang Zeng
, Yunkun Shi
, Kai Liu^*

^*Corresponding author for this work

XJTLU

Research output: Contribution to conference › Paper › peer-review

3 Downloads (Pure)

Abstract

Federated credit risk prediction can reduce pooling of client training records, but label imbalance, client heterogeneity, and private training noise impair utility. HA-DP-FedAvg combines an RDP-calibrated back-loaded noise schedule with aggregation weights derived from one-time private, shrinkage-stabilized label prevalence. We evaluate three credit datasets under moderate and severe Dirichlet skew using 50 rounds, total ϵ ≈ 3, and ten paired seeds. Relative to privacy-matched DP-FedAvg, mean AUPRC increases from 0.422 to 0.460 on Default Credit and from 0.330 to 0.369 on Give Me Some Credit at moderate skew; at severe skew, the changes are 0.431 to 0.442 and 0.289 to 0.319. German Credit results are smaller and inconclusive. Factorial and horizon analyses show context-dependent gains that narrow with longer training.

Original language	English
Publication status	Published - 27 Nov 2026
Event	2026 IEEE 9th International Conference on Data Science and Information Technology - Xi'an Jiaotong-Liverpool University, Taicang, China Duration: 27 Nov 2026 → 29 Nov 2026

Conference

Conference	2026 IEEE 9th International Conference on Data Science and Information Technology
Abbreviated title	DSIT 2026
Country/Territory	China
City	Taicang
Period	27/11/26 → 29/11/26

Keywords

Federated learning, Differential privacy, Credit risk, Class imbalance, Non-IID data, Privacy accounting

Access to Document

HA-DP-FedAvg: Privacy Scheduling and Heterogeneity-Aware Aggregation for Federated Credit Risk PredictionFinal published version, 4.61 MB

Cite this

@conference{356c3c710d324133bc8ceb692535bfd5,

title = "HA-DP-FedAvg: Privacy Scheduling and Heterogeneity-Aware Aggregation for Federated Credit Risk Prediction",

abstract = "Federated credit risk prediction can reduce pooling of client training records, but label imbalance, client heterogeneity, and private training noise impair utility. HA-DP-FedAvg combines an RDP-calibrated back-loaded noise schedule with aggregation weights derived from one-time private, shrinkage-stabilized label prevalence. We evaluate three credit datasets under moderate and severe Dirichlet skew using 50 rounds, total ϵ ≈ 3, and ten paired seeds. Relative to privacy-matched DP-FedAvg, mean AUPRC increases from 0.422 to 0.460 on Default Credit and from 0.330 to 0.369 on Give Me Some Credit at moderate skew; at severe skew, the changes are 0.431 to 0.442 and 0.289 to 0.319. German Credit results are smaller and inconclusive. Factorial and horizon analyses show context-dependent gains that narrow with longer training.",

keywords = "Federated learning, Differential privacy, Credit risk, Class imbalance, Non-IID data, Privacy accounting",

author = "Zihang Zeng and Yunkun Shi and Kai Liu",

year = "2026",

month = nov,

day = "27",

language = "English",

note = "2026 IEEE 9th International Conference on Data Science and Information Technology, DSIT 2026 ; Conference date: 27-11-2026 Through 29-11-2026",

}

TY - CONF

T1 - HA-DP-FedAvg

T2 - 2026 IEEE 9th International Conference on Data Science and Information Technology

AU - Zeng, Zihang

AU - Shi, Yunkun

AU - Liu, Kai

PY - 2026/11/27

Y1 - 2026/11/27

N2 - Federated credit risk prediction can reduce pooling of client training records, but label imbalance, client heterogeneity, and private training noise impair utility. HA-DP-FedAvg combines an RDP-calibrated back-loaded noise schedule with aggregation weights derived from one-time private, shrinkage-stabilized label prevalence. We evaluate three credit datasets under moderate and severe Dirichlet skew using 50 rounds, total ϵ ≈ 3, and ten paired seeds. Relative to privacy-matched DP-FedAvg, mean AUPRC increases from 0.422 to 0.460 on Default Credit and from 0.330 to 0.369 on Give Me Some Credit at moderate skew; at severe skew, the changes are 0.431 to 0.442 and 0.289 to 0.319. German Credit results are smaller and inconclusive. Factorial and horizon analyses show context-dependent gains that narrow with longer training.

AB - Federated credit risk prediction can reduce pooling of client training records, but label imbalance, client heterogeneity, and private training noise impair utility. HA-DP-FedAvg combines an RDP-calibrated back-loaded noise schedule with aggregation weights derived from one-time private, shrinkage-stabilized label prevalence. We evaluate three credit datasets under moderate and severe Dirichlet skew using 50 rounds, total ϵ ≈ 3, and ten paired seeds. Relative to privacy-matched DP-FedAvg, mean AUPRC increases from 0.422 to 0.460 on Default Credit and from 0.330 to 0.369 on Give Me Some Credit at moderate skew; at severe skew, the changes are 0.431 to 0.442 and 0.289 to 0.319. German Credit results are smaller and inconclusive. Factorial and horizon analyses show context-dependent gains that narrow with longer training.

KW - Federated learning, Differential privacy, Credit risk, Class imbalance, Non-IID data, Privacy accounting

M3 - Paper

Y2 - 27 November 2026 through 29 November 2026

ER -