OPUS-VFL: Incentivizing Optimal Privacy-Utility Tradeoffs in Vertical Federated Learning

Sindhuja Madabushi, Ahmad Faraz Khan, Haider Ali, Jin-Hee Cho

Abstract

Vertical Federated Learning (VFL) enables organizations with disjoint feature spaces but shared user bases to collaboratively train models without sharing raw data. However, existing VFL systems face critical limitations: they often lack effective incentive mechanisms, struggle to balance privacy-utility tradeoffs, and fail to accommodate clients with heterogeneous resource capabilities. These challenges hinder meaningful participation, degrade model performance, and limit practical deployment. To address these issues, we propose OPUS-VFL, an Optimal Privacy-Utility tradeoff Strategy for VFL. OPUS-VFL introduces a novel, privacy-aware incentive mechanism that rewards clients based on a principled combination of model contribution, privacy preservation, and resource investment. It employs a lightweight leave-one-out (LOO) strategy to quantify feature importance per client, and integrates an adaptive differential privacy mechanism that enables clients to dynamically calibrate noise levels to optimize their individual utility. Our framework is designed to be scalable, budget-balanced, and robust to inference and poisoning attacks. Extensive experiments on benchmark datasets (MNIST, CIFAR-10, and CIFAR-100) demonstrate that OPUS-VFL significantly outperforms state-of-the-art VFL baselines in both efficiency and robustness. It reduces label inference attack success rates by up to 20%, increases feature inference reconstruction error (MSE) by over 30%, and achieves up to 25% higher incentives for clients that contribute meaningfully while respecting privacy and cost constraints. These results highlight the practicality and innovation of OPUS-VFL as a secure, fair, and performance-driven solution for real-world VFL.

Introduction

In Vertical Federated Learning (VFL), different organizations (such as a bank and a retail chain) hold different sets of features for the same group of users. While this allows for powerful collaborative models, it faces three major hurdles:

  1. Incentive Gaps: Participants lack motivation to expend computational resources and risk data exposure without fair rewards.
  2. The Privacy-Utility Tightrope: Traditional Differential Privacy (DP) often degrades model performance significantly, making the optimal noise level difficult to determine.
  3. Client Heterogeneity: Systems often fail to account for varying hardware capabilities and data quality across clients.
Comparison of Horizontal and Vertical Federated Learning architectures.

Without a mechanism ensuring economic fairness and robust privacy, VFL adoption remains limited in sensitive sectors like healthcare and finance. OPUS-VFL addresses these by aligning economic incentives with privacy and utility.

Methodology

OPUS-VFL (Optimal Privacy-Utility tradeoff Strategy) operates on a Split Neural Network (SplitNN) architecture. Its core technical components include:

Lightweight Contribution Evaluation

Instead of computationally expensive methods like Shapley values, OPUS-VFL utilizes a Leave-One-Out (LOO) strategy. The server calculates the global loss with and without a specific client’s features to determine their marginal contribution efficiently.

Adaptive Differential Privacy (DP)

Clients inject Gaussian noise into local model updates before transmission. Crucially, clients use a gradient-based approach to dynamically tune their privacy parameter ($\epsilon$) based on received rewards. This treats the reward signal as a feedback loop, allowing clients to find the $\epsilon$ that maximizes their individual utility.

Resource-Sensitive Incentive Design

The reward mechanism factors in "contribution equity," rewarding clients based on model impact, privacy-preservation efforts, and computational resource investment. This ensures the system is budget-balanced and "individually rational," meaning participation remains profitable for the client.

Experimental Evaluation

The framework was evaluated using standard machine learning benchmarks: MNIST, CIFAR-10, and CIFAR-100.

Key Results

  • Defense Against Attacks: OPUS-VFL reduced label inference attack success rates by up to 20% compared to vanilla VFL. For feature inference attacks, it increased the reconstruction error (MSE) by over 30%.
  • Efficiency: The system achieved a 20.8% reduction in computation time compared to the "Bid Price First" (BPF) baseline and was 88% more efficient than VFL-CZOFO.
  • Fairer Incentives: High-quality, privacy-respecting contributors received up to 25% higher incentives than in traditional baseline models.
  • Scalability: Testing with up to 25 clients on MNIST showed that the system maintains competitive accuracy (peaking at 92.5%) while training time scales linearly.

Conclusion

OPUS-VFL provides a secure, fair, and performance-driven solution for real-world VFL. By integrating an adaptive privacy mechanism with a resource-sensitive incentive design, it overcomes the traditional barriers to collaborative learning between organizations. Future work will explore mechanisms to verify client-reported costs and improve robustness against network stragglers.

Summary updated

Abstract

Vertical Federated Learning (VFL) enables organizations with disjoint feature spaces but shared user bases to collaboratively train models without sharing raw data. However, existing VFL systems face critical limitations: they often lack effective incentive mechanisms, struggle to balance privacy-utility tradeoffs, and fail to accommodate clients with heterogeneous resource capabilities. These challenges hinder meaningful participation, degrade model performance, and limit practical deployment. To address these issues, we propose OPUS-VFL, an Optimal Privacy-Utility tradeoff Strategy for VFL. OPUS-VFL introduces a novel, privacy-aware incentive mechanism that rewards clients based on a principled combination of model contribution, privacy preservation, and resource investment. It employs a lightweight leave-one-out (LOO) strategy to quantify feature importance per client, and integrates an adaptive differential privacy mechanism that enables clients to dynamically calibrate noise levels to optimize their individual utility. Our framework is designed to be scalable, budget-balanced, and robust to inference and poisoning attacks. Extensive experiments on benchmark datasets (MNIST, CIFAR-10, and CIFAR-100) demonstrate that OPUS-VFL significantly outperforms state-of-the-art VFL baselines in both efficiency and robustness. It reduces label inference attack success rates by up to 20%, increases feature inference reconstruction error (MSE) by over 30%, and achieves up to 25% higher incentives for clients that contribute meaningfully while respecting privacy and cost constraints. These results highlight the practicality and innovation of OPUS-VFL as a secure, fair, and performance-driven solution for real-world VFL.
BibTeX
@misc{madabushi2025opus,
  author = {Sindhuja Madabushi and Ahmad Faraz Khan and Haider Ali and Jin-Hee Cho},
  title = {{OPUS-VFL: Incentivizing Optimal Privacy-Utility Tradeoffs in Vertical Federated Learning}},
  year = {2025},
  eprint = {2504.15995},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.15995},
}
Figure