Get Started
Home
Topics
Search
Library
Research questionHow can we reliably detect memorized patient images in medical generative models at scale?Medical image generators may reproduce or closely resemble training patients’ images, creating confidentiality risks. Pixel comparisons can be distorted by generation artifacts, while generic embeddings may miss anatomically meaningful similarity.
AI
AI Memory
Alignment & Safety
Computer Vision
Evaluation & Benchmarks
Health
Image & Video Processing
Image Generation
Machine Learning
Latest papersRecent research connected to this question, newest first.Auditing Patient Privacy in Medical Generative Models: Scalable Memorization Detection with DeepSSIM++The source concerns scalable memorization auditing for medical generative models using generated and training medical images. It reports DeepSSIM++, a self-supervised similarity metric with multi-scale features and anatomy-preserving augmentations, evaluated under ideal alignment and realistic spatial and intensity perturbations against existing baselines. The reported evidence also covers faster similarity computation than analytical SSIM, with code and data publicly available.research paper · Sep 3, 2026
Related questions
How can open image generators follow fine-grained editing instructions without sacrificing image quality or inference speed?How can we compare medical VQA models when generation obscures answer signal and biases answer positions?How can medical-image counterfactuals reveal a classifier’s decisions without inheriting a generator’s biases?How can histopathology models withstand adversarial perturbations without erasing diagnostic tissue structures?