Privacy-Preserving Data Synthetisation for Secure Information Sharing

Tânia Carvalho, Nuno Moniz, Pedro Faria, Luís Antunes, Nitesh V. Chawla

Abstract

Protecting user data privacy can be achieved via many methods, from statistical transformations to generative models. However, all of them have critical drawbacks. For example, creating a transformed data set using traditional techniques is highly time-consuming. Also, recent deep learning-based solutions require significant computational resources in addition to long training phases, and differentially private-based solutions may undermine data utility. In this paper, we propose $\epsilon$-PrivateSMOTE, a technique designed for safeguarding against re-identification and linkage attacks, particularly addressing cases with a high re-identification risk. Our proposal combines synthetic data generation via noise-induced interpolation with differential privacy principles to obfuscate high-risk cases. We demonstrate how $\epsilon$-PrivateSMOTE is capable of achieving competitive results in privacy risk and better predictive performance when compared to multiple traditional and state-of-the-art privacy-preservation methods, including generative adversarial networks, variational autoencoders, and differential privacy baselines. We also show how our method improves time requirements by at least a factor of 9 and is a resource-efficient solution that ensures high performance without specialised hardware.

Introduction

Data sharing is the lifeblood of modern research and machine learning, yet it remains a minefield for privacy. Traditional "de-identification" methods, such as generalization (grouping ages into ranges) or suppression (deleting values), often strip away the very nuances that make datasets valuable for predictive modeling.

On the other end of the spectrum, modern generative models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) can create synthetic data from scratch. However, these models are computationally expensive, require massive amounts of data to learn distributions effectively, and often struggle to protect outliers—the individuals most at risk of being "singled out" in a re-identification attack. This research addresses the critical gap for a method that protects high-risk individuals without destroying the utility of the entire dataset or requiring specialized high-performance computing resources.

The $\epsilon$-PrivateSMOTE Framework

The proposed $\epsilon$-PrivateSMOTE algorithm introduces a targeted approach to privacy. Instead of transforming the entire dataset, it identifies specific high-risk records and replaces them with synthetic versions created through differentially private interpolation. The framework operates in three primary stages:

  1. Highest-Risk Selection: The algorithm uses $k$-anonymity principles to identify "high-risk" records—those that are unique or nearly unique based on their quasi-identifiers (such as age, gender, and zip code). By focusing only on these records, the method avoids over-anonymizing "safe" data.
  2. SMOTE-Inspired Interpolation: For these high-risk cases, the method generates synthetic replacements. It identifies the nearest neighbors of a record and creates a new point through interpolation. While this is a standard technique in imbalanced learning (SMOTE), it is repurposed here to mask the identity of sensitive records.
  3. Differential Privacy (DP) Integration: To ensure these new points do not inadvertently leak information, the authors incorporate the Laplace mechanism. Standard uniform noise is replaced with noise drawn from a Laplace distribution, parameterized by a privacy budget ($\epsilon$). This ensures that the resulting synthetic data is statistically indistinguishable from the original sensitive points.

Experimental Setup

The researchers evaluated $\epsilon$-PrivateSMOTE using 15 diverse classification datasets sourced from the OpenML repository. These datasets cover various real-world domains with binary class targets, ranging from approximately 1,000 to 15,000 instances and feature counts between 5 and 103. On average, the datasets contained 28 numerical attributes and 2 nominal attributes, providing a robust testing ground for tabular data synthesis.

Key Results and Insights

Predictive Performance and Utility

In terms of Area Under the ROC Curve (AUC), $\epsilon$-PrivateSMOTE demonstrated a 60% probability of practical equivalence to the "oracle" (the best model trained on original, unprotected data). This indicates that the method maintains high data utility, matching the performance of traditional techniques while offering significantly stronger privacy guarantees.

Privacy Risk Mitigation

The method achieved a near-zero linkability risk, which measures the ability of an attacker to link synthetic records back to real individuals. This performance significantly outperformed standard sampling methods like regular SMOTE or Random Under-Sampling, which do not provide formal privacy protections.

Computational Efficiency

Efficiency is a standout result of this research. $\epsilon$-PrivateSMOTE was found to be at least 9 times faster than PATE-GAN. In the time it took a GAN-based approach to generate only 240 data variants, $\epsilon$-PrivateSMOTE generated over 3,300. Furthermore, because the method is a targeted heuristic rather than a deep learning model, it does not require a GPU. Running the algorithm on a CPU was actually more efficient, as it avoided the latency associated with moving data between processors.

Conclusion

$\epsilon$-PrivateSMOTE demonstrates that targeted protection is often superior to global blurring. By focusing its "blurring" power only where it is needed most—on the high-risk tails of the distribution—the method preserves the core utility of the dataset. While currently limited to tabular data, the approach offers a resource-efficient solution for organizations that need to share sensitive information securely without the overhead of complex generative models. Future work will focus on expanding the method to handle more complex inference-based attacks and investigating the persistence of outliers in synthetic distributions.

Summary updated

Abstract

Protecting user data privacy can be achieved via many methods, from statistical transformations to generative models. However, all of them have critical drawbacks. For example, creating a transformed data set using traditional techniques is highly time-consuming. Also, recent deep learning-based solutions require significant computational resources in addition to long training phases, and differentially private-based solutions may undermine data utility. In this paper, we propose $\epsilon$-PrivateSMOTE, a technique designed for safeguarding against re-identification and linkage attacks, particularly addressing cases with a high \sloppy re-identification risk. Our proposal combines synthetic data generation via noise-induced interpolation with differential privacy principles to obfuscate high-risk cases. We demonstrate how $\epsilon$-PrivateSMOTE is capable of achieving competitive results in privacy risk and better predictive performance when compared to multiple traditional and state-of-the-art privacy-preservation methods, including generative adversarial networks, variational autoencoders, and differential privacy baselines. We also show how our method improves time requirements by at least a factor of 9 and is a resource-efficient solution that ensures high performance without specialised hardware.
BibTeX
@article{carvalho2025privacy,
  author = {Tânia Carvalho and Nuno Moniz and Pedro Faria and Luís Antunes and Nitesh V. Chawla},
  title = {{Privacy-Preserving Data Synthetisation for Secure Information Sharing}},
  journal = {Machine Learning},
  year = {2025},
  doi = {10.1007/s10994-025-06799-w},
}
Figure