Privacy-Enhanced Vertical Federated Learning for Healthcare via Directional Noise and Subset Representations.

Abstract

Vertical federated learning (VFL) allows healthcare institutions to train models on complementary patient features without sharing raw data, but strong differential privacy often causes severe utility loss and labeled medical data are limited.We propose HEAL, a privacy-enhanced VFL framework that jointly learns subset representations and optimizes the direction of privacy-preserving noise. HEAL first constructs importance-aware feature subsets and performs multi-level contrastive pre-training to exploit unlabeled data and unify heterogeneous feature spaces. It then applies direction-optimized differential privacy to preserve formal $(\epsilon, \delta)$-privacy while reducing gradient distortion, followed by collaborative task learning for healthcare prediction. Across four healthcare datasets, HEAL improves accuracy by 2.6-4.7% over state-of-the-art baselines, reaches 96.2% of centralized performance at $\epsilon =1.0$, and degrades gradient-inversion reconstruction quality by 20-35%. These results show that privacy protection and representation learning can reinforce each other, rather than treating privacy only as a performance cost.