Unconditional latent diffusion models memorize patient imaging data.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Data Privacy

Background:

• Generative AI models create synthetic data for open sharing, but risk patient data memorization and re-identification.
• Patient data memorization occurs when models replicate, rather than generate novel, samples.

Purpose of the Study:

• To assess data memorization in unconditional latent diffusion models.
• To evaluate the susceptibility of generative models to patient data memorization.

Main Methods:

• Trained unconditional latent diffusion models on diverse datasets.
• Employed a self-supervised copy detection approach to identify memorized patient data.
• Compared memorization rates across different generative models (diffusion vs. autoencoders, GANs).

Main Results:

• High patient data memorization observed across all datasets (37.2% memorized, 68.7% synthetic samples as copies).
• Latent diffusion models exhibited greater memorization than autoencoders and GANs, despite superior synthesis quality.
• Training strategies like augmentation, smaller architectures, and larger datasets reduced memorization; overtraining increased it.

Conclusions:

• Latent diffusion models pose significant privacy risks due to high data memorization.
• Careful training and validation are crucial when using generative models for private medical imaging data.
• Ensuring synthetic data integrity is vital for maintaining patient privacy in AI-driven healthcare.