Enabling PSO-Secure Synthetic Data Sharing Using Diversity-Aware Diffusion Models

Mischa Dombrowski, Bernhard Kainz

Introduction

Sharing medical data for AI research is a complex challenge governed by strict legal frameworks like the General Data Protection Regulation (GDPR). While synthetic data—images generated by AI rather than sourced from real patients—offers a potential path forward, it has historically struggled with two primary issues: legal anonymity and diagnostic utility.

This paper addresses the risk of Predicate Singling-Out (PSO), a scenario where a combination of unique traits (predicates) allows an individual to be identified within a dataset. Furthermore, it tackles the "utility gap," where models trained on synthetic data often underperform compared to those trained on real-world clinical data. The researchers propose a framework using Diversity-Aware Diffusion Models (DiADM) to generate synthetic medical images that are both PSO-secure and high-performing.

Methodology

The proposed framework adapts the Diversity-Aware Diffusion Model (DiADM) and integrates a multi-stage privacy and utility pipeline:

  1. Feature Extraction via SwAV: Instead of relying on simple diagnostic labels, the model uses a pre-trained SwAV (Self-supervised Learning of Visual Features) model to extract complex visual predicates from real images. These predicates represent the clinical essence of the image without capturing specific patient identities.
  2. Diversity-Aware Generation: The model generates multiple variations of an image based on these predicates. By ensuring that many different synthetic images can represent the same clinical traits, the system prevents any single image from being uniquely linked to a training sample.
  3. Privacy Filtering: To ensure PSO-security, every generated image is passed through a re-identification model. If a synthetic image is found to be too similar to a real training sample (suggesting "memorization"), it is discarded to prevent identity leaks.
  4. Predicate Alignment (BCE Ranking): The remaining images are ranked using a classifier trained on real data. This ensures that the synthetic images preserve the critical clinical features necessary for diagnostic tasks.

Experimental Setup

The study utilized three major chest X-ray datasets: MIMIC-CXR, CheXpert (CXP), and ChestX-ray8 (NIH). The researchers focused on eight shared disease classes and employed a 70/10/20 split for training, validation, and testing. To ensure clean benchmarking against state-of-the-art (SOTA) methods, the team specifically filtered for images containing single pathology labels.

Key Results and Insights

The DiADM + SwAV method demonstrated significant improvements over existing generative models like EDM-2:

  • Fidelity and Diversity: The model achieved a Fréchet Inception Distance (FID) of 5.0, a marked improvement over the 8.9–15.0 range seen in competing models.
  • Image Retrieval Score (IRS): With an IRS of 1.58, the model generated more diverse samples than even a theoretical "perfect" unconditional model, supporting the hypothesis that diversity enhances privacy.
  • Diagnostic Utility: On the NIH dataset, the model achieved an AUCROC of 84.04%. This is remarkably close to the 84.99% achieved by models trained on real patient data, effectively closing the utility gap to less than 1%.
  • Generalization: Models trained on this synthetic data showed robust performance when tested on entirely different real-world datasets, consistently outperforming other synthetic generators in cross-domain tasks.

Discussion: Diversity as Privacy

A core insight of this research is that maximizing image diversity serves as a privacy mechanism. By ensuring that a wide variety of synthetic representations exist for any given set of clinical traits, the likelihood of a single image being traced back to a specific individual is minimized.

Interestingly, the researchers found that using a general-purpose feature extractor (SwAV) outperformed domain-specific medical models. They hypothesize that domain-specific features may be "too meaningful," leading to reconstruction losses that hinder the diffusion model's ability to generate high-quality variations.

Limitations and Future Work

Despite these advancements, a small performance gap remains between synthetic and real data. Additionally, the computational cost of the sampling process is high, as the model must generate and filter a large batch of images to select the best candidates. The security of the system also remains dependent on the efficacy of the re-identification filter.

Future research will explore the use of these synthetic datasets for data augmentation, particularly in low-resource medical settings where real data is scarce, to see if synthetic data can eventually help AI models surpass real-data performance ceilings.

Summary updated

BibTeX
@inproceedings{dombrowski2025enabling,
  author = {Mischa Dombrowski and Bernhard Kainz},
  title = {{Enabling PSO-Secure Synthetic Data Sharing Using Diversity-Aware Diffusion Models}},
  booktitle = {BRIDGE/DeCaF @ MICCAI},
  year = {2025},
  doi = {10.1007/978-3-032-05663-4\_3},
}
Figure