Privacy in Image Datasets: A Case Study on Pregnancy Ultrasounds

Large-scale datasets contain sensitive medical information: The researchers identified thousands of pregnancy ultrasound images in LAION-400M, a widely-used dataset for training generative models, demonstrating that sensitive personal health data is being collected and used without explicit consent.

Private identifying information is readily accessible: The study detected thousands of instances of private information (names, locations, and other personal details) embedded within or associated with ultrasound images in the dataset, creating significant re-identification and impersonation risks for individuals.

Current data collection practices lack adequate safeguards: The paper highlights that large-scale internet datasets are typically collected "with minimal or no data curation," meaning sensitive medical images can be included without review or filtering for privacy concerns.

Generative AI amplifies privacy risks: As these datasets are used to train generative models, the private information they contain could potentially be reproduced or leaked through model outputs, extending privacy violations beyond the original dataset.

Dataset curation and ethical guidelines are needed: The authors recommend implementing better practices for dataset curation, data privacy protections, and ethical standards for public image dataset use to prevent sensitive personal information from being included in AI training data.