InC: A Vertical Federated Learning Framework with Multiple Noisy Labels

Ruixuan Zhang, Ziyi Li, Xiao Yan, Xiaokai Zhou, Hao Wang, Hao Huang, Jiawei Jiang

Abstract: Vertical Federated Learning (VFL) enables collaborative model training in a privacy-preserving way when data features are distributed across different parties. Typical VFL systems assume only one party holds the training labels. However, in real-world scenarios, each party may annotate its local data independently, resulting in significant label noise due to incomplete feature spaces and varying annotation quality. To address this challenge, we propose Initialization Correction (InC), a VFL framework designed to handle multiple noisy labels. In the first stage, we train a consensus classifier by converting party labels into probability distributions to enhance noise tolerance. In the second stage, a temporary corrected label is inferred and iteratively refined through an EM process. The expertise matrix of each party is evaluated and used to weight the classifier predictions, which serve as the corrected label. Throughout the entire process, we leverage techniques such as Homomorphic Encryption (HE) to protect the original labels’ privacy. Extensive experiments on public datasets demonstrate that InC significantly outperforms the baselines across various settings, and the results are comparable to training with clean labels.

External IDs:dblp:conf/dasfaa/ZhangLYZWHJ25