A Quantitative Metric for Privacy Leakage in Federated Learning

Yong Liu, Xinghua Zhu, Jianzong Wang, Jing Xiao

Abstract

In Federated Learning (FL), while raw data remains on client devices, the sharing of parameter gradients can inadvertently expose sensitive information to inference attacks. This paper proposes a quantitative metric based on Mutual Information (MI) to evaluate the risk of information leakage. To overcome the limitations of existing MI estimation methods in high-dimensional spaces, we introduce H-MINE (Hierarchical Mutual Information Neural Estimation). Our method reliably reflects information leakage levels and identifies key influential factors, such as model training status and data distribution. Experimental results on the Adult dataset demonstrate that H-MINE provides a robust "risk pre-alarm" for FL clients.

Introduction

Federated Learning is widely regarded as a privacy-preserving paradigm because it avoids the direct exchange of raw data. However, recent studies have demonstrated that an adversary—including a curious central server—can reconstruct original data by analyzing intercepted gradients. While defenses like Differential Privacy (DP) exist, they often lack a clear, interpretable metric for the actual privacy risk incurred during a specific update.

This paper identifies Mutual Information (MI) as a universal, quantifiable metric for this risk. However, estimating MI for high-dimensional variables (like gradients in modern neural networks) is computationally challenging. We address this by proposing H-MINE, a hierarchical approach designed to handle high-dimensional batch data efficiently.

System Model and Risk Detection

The proposed framework integrates a local risk detection module within the standard FL workflow. Before a client transmits a gradient update to the central server, it utilizes the MI estimator to calculate the potential leakage risk.

The proposed federated learning system model with integrated local risk detection.

As shown in the system model, if the estimated MI exceeds a predefined safety threshold, the client can choose to discard the update or apply additional privacy-preserving measures (e.g., noise injection). This "Risk Pre-Alarm" empowers clients to make informed decisions about their data privacy before any information leaves their local environment.

Methodology: H-MINE Architecture

The core innovation of this work is the Hierarchical Mutual Information Neural Estimation (H-MINE). Standard MI estimators often fail to converge when dealing with high-dimensional inputs. H-MINE overcomes this by exploiting the Independent and Identically Distributed (IID) nature of batch data through a two-stage architecture.

Architecture of the H-MINE estimator, featuring BlockModel and MixModel components.
  1. BlockModel: A shared sub-network processes individual data-gradient pairs $(x_i, G)$ into lower-dimensional embedding vectors $h_i$. This significantly reduces the parameter space compared to processing the entire batch at once.
  2. MixModel: These embeddings are concatenated and passed through a secondary network to produce a final scalar value $T(X, G)$, which is used to approximate the MI based on the Donsker-Varadhan representation of KL divergence.

Experimental Evaluation and Insights

We evaluated H-MINE using the Adult dataset, performing a binary classification task (income prediction). The global model was a logistic regression, and we measured the MI between the input batch and the resulting gradients.

Key Findings

  • Convergence and Scalability: Unlike standard MINE, which often diverges in high-dimensional settings, H-MINE demonstrates stable convergence even as batch sizes and dimensions increase.
  • The "Early Bird" Risk: Privacy leakage is not uniform throughout the training process. Leakage risk is highest during the initial epochs when the model is most sensitive to raw data. As the model converges, the gradients carry less novel information, and the MI stabilizes.
  • Data Distribution Impact: The inherent distribution of the data significantly affects privacy. Experiments show that highly unbalanced datasets (e.g., extreme ratios of positive to negative samples) are more susceptible to information leakage through gradients.
  • Correlation with Attack Success: The paper validates that the MI metric correlates strongly with the success of actual inference attacks (e.g., DeepLeakage). A higher estimated MI corresponds to a lower reconstruction error for an adversary.

Conclusion

This research provides a practical, quantifiable metric for privacy in Federated Learning. By introducing H-MINE, we enable clients to estimate the information leakage risk of high-dimensional gradients locally. This metric not only justifies common heuristics like batch size increases and noise addition but also provides a mathematical foundation for developing more robust privacy-preserving FL protocols. Future work will focus on scaling H-MINE to larger deep learning architectures and exploring its performance in non-IID data environments.

Summary updated

Abstract

In the federated learning system, parameter gradients are shared among participants and the central modulator, while the original data never leave their protected source domain. However, the gradient itself might carry enough information for precise inference of the original data. By reporting their parameter gradients to the central server, client datasets are exposed to inference attacks from adversaries. In this paper, we propose a quantitative metric based on mutual information for clients to evaluate the potential risk of information leakage in their gradients. Mutual information has received increasing attention in the machine learning and data mining community over the past few years. However, existing mutual information estimation methods cannot handle high-dimensional variables. In this paper, we propose a novel method to approximate the mutual information between the high-dimensional gradients and batched input data. Experimental results show that the proposed metric reliably reflect the extent of information leakage in federated learning. In addition, using the proposed metric, we investigate the influential factors of risk level. It is proven that, the risk of information leakage is related to the status of the task model, as well as the inherent data distribution.
BibTeX
@inproceedings{liu2021quantitative,
  author = {Yong Liu and Xinghua Zhu and Jianzong Wang and Jing Xiao},
  title = {{A Quantitative Metric for Privacy Leakage in Federated Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2021},
  doi = {10.1109/icassp39728.2021.9413539},
}
Figure