Regression Model Fitting under Differential Privacy and Model Inversion Attack

Yue Wang, Cheng Si, Xintao WuIJCAI 2015

Abstract: Standard Differential Privacy (DP) provides strong guarantees for membership privacy but often leaves sensitive individual attributes vulnerable to model inversion attacks. This paper proposes a modified Functional Mechanism that employs a non-uniform noise distribution strategy. By strategically allocating a larger portion of the privacy budget to non-sensitive attributes and heavily perturbing terms associated with sensitive features, the proposed method effectively thwarts inversion attacks while preserving the overall utility of regression models. Empirical evaluations on the Adult dataset show that this targeted perturbation approach outperforms uniform noise allocation in balancing the privacy-utility trade-off.

Introduction

In privacy-preserving data mining, Differential Privacy (DP) ensures that the inclusion of any single individual's data does not significantly alter the output of a learning algorithm. However, a critical vulnerability remains: attribute privacy. Even with DP guarantees, an adversary with background knowledge can perform a model inversion attack. By inputting known non-sensitive data into a public model, the attacker can infer sensitive attributes, such as medical conditions or financial status. Previous attempts to mitigate these attacks often resulted in a prohibitive loss of model accuracy. This research addresses this gap by proposing a method to protect specific attributes without compromising the model's overall predictive power.

Methodology

The proposed approach builds on the Functional Mechanism, which achieves DP by adding noise to the objective function during training. The core innovation is a non-uniform noise distribution strategy:

  1. Polynomial Representation: The regression objective function is represented as a polynomial.
  2. Attribute Splitting: Features are divided into sensitive ($X_s$) and non-sensitive ($X_n$) sets.
  3. Budget Balancing: A ratio parameter $\gamma$ is introduced to allocate the total privacy budget $\epsilon$. A smaller portion of the budget is assigned to terms involving sensitive attributes, leading to higher Laplace noise in those specific areas.
Table 1: Allocation of privacy budgets for sensitive ($\epsilon_s$) and non-sensitive ($\epsilon_n$) attributes based on the ratio parameter $\gamma$.

This targeted perturbation "blurs" the mathematical relationship between sensitive features and the model output, making it significantly harder for an attacker to invert the logic.

Experimental Evaluation

The method was evaluated using the Adult Dataset (UCI Machine Learning Repository), with the task of predicting whether an individual's income exceeds $50,000. "Marital status" was designated as the sensitive attribute for the model inversion attack experiments.

Key Results

The experiments highlight a granular utility-privacy trade-off. As the budget ratio $\gamma$ decreases, the accuracy of the model inversion attack drops sharply toward the baseline of random guessing.

Figure 1: Impact of the budget ratio $\gamma$ on Model Accuracy and Attack Accuracy. As $\gamma$ decreases (moving right), attack accuracy drops significantly while model utility remains high.

As shown in Figure 1, when $\gamma$ is set to 0.01, attack accuracy falls to approximately 57% (near the 53% baseline), while the overall model accuracy remains robust, staying above 75%. This demonstrates that strategic noise allocation can protect sensitive attributes more effectively than uniform noise, which would require a much lower total budget $\epsilon$ and thus a greater loss in utility to achieve the same level of protection.

Conclusion

This work demonstrates that membership privacy and attribute privacy are distinct challenges. By identifying and aggressively protecting the "leaky" channels associated with sensitive attributes, researchers can build models that are resilient to inversion attacks without sacrificing general performance. Future work will explore the application of this technique to more complex data types, such as genomic data, and scenarios involving multiple sensitive attributes.

Summary updated

BibTeX
@inproceedings{wang2015regression,
  author = {Yue Wang and Cheng Si and Xintao Wu},
  title = {{Regression Model Fitting under Differential Privacy and Model Inversion Attack}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure