Abstract The rise of generative models has led to increased use of large-scale datasets collected from the internet, often with minimal or no data curation. This raises concerns about the inclusion of sensitive or private information. In this work, we explore the presence of pregnancy ultrasound images, which contain sensitive personal information and are often shared online. Through a systematic examination of LAION-400M dataset using CLIP embedding similarity, we retrieve images containing pregnancy ultrasound and detect thousands of entities of private information such as names and locations. Our findings reveal that multiple images have high-risk information that could enable re-identification or impersonation. We conclude with recommended practices for dataset curation, data privacy, and ethical use of public image datasets.
Introduction: The Privacy Crisis in Large-Scale Datasets
Modern computer vision and generative AI, such as Stable Diffusion, are built upon the foundation of massive datasets like LAION. These datasets are typically compiled by indiscriminately scraping billions of images and captions from the public web. While this scale enables unprecedented model performance, it creates a "wild west" environment where sensitive, non-consensual data is ingested into training pipelines.
This study focuses on pregnancy ultrasound images as a primary case study for privacy violations. Ultrasounds represent a unique intersection of clinical medical records and social media content. While they are private medical documents, expectant parents frequently share them online for announcements or invitations. This dual nature makes them highly susceptible to being scraped into AI training sets, potentially allowing models to memorize and reproduce specific personal identifiable information (PII).
Methodology: Detection and Extraction Pipelines
To audit the LAION-400M dataset, researchers developed a multi-stage pipeline designed to first identify relevant images and then extract embedded text.
Detection Strategies
The team compared two distinct approaches to locate ultrasounds within the 400 million images:
- Retrieval-based: Utilizing CLIP (Contrastive Language-Image Pre-training) embeddings to find images that matched specific text queries (e.g., "pregnancy announcement") or visual examples.
- Classifier-based: Training traditional machine learning models, such as Support Vector Machines (SVMs) and Random Forests, on a custom-curated PIU (Pregnancy Image Ultrasound) Dataset. This dataset contains 9,900 images, balanced between actual ultrasounds and "hard negatives"—images that look similar to ultrasounds but are not (e.g., X-rays, certain animals, or household objects).
Surprisingly, the SVM outperformed complex neural networks in this specific task, achieving 97.27% accuracy. This high precision is vital for large-scale audits where false positives can lead to massive amounts of irrelevant data.
PII Extraction
Once the ultrasounds were identified, a sophisticated OCR (Optical Character Recognition) pipeline was deployed:
- Preprocessing: Images were upscaled using Real-ESRGAN and rotated to ensure text was horizontal.
- OCR & Vision-Language Correction: Initial text was extracted via Tesseract. To handle the low resolution and noise typical of medical scans, the researchers used LLaVA-Next, a vision-language model, to correct OCR errors by cross-referencing the text with the visual context of the image.
- Entity Recognition: Microsoft’s Presidio was used to categorize the cleaned text into four PII categories: Name, Location, Date/Time, and Phone Number.
The Social Context of Medical Data
A key insight from the audit is the concept of the "Social Trojan Horse." Many of the sensitive images found were not direct clinical uploads but were embedded within social artifacts like baby shower invitations or pregnancy announcements.
These social contexts act as PII goldmines. While a clinical scan might only have a hospital name, an invitation often includes full names, home addresses, and specific event dates. This transition of data from a private social circle to a global AI training set violates the principle of contextual integrity, where privacy is breached because the data is used in a context entirely different from the one intended by the subject.
Key Results and Findings
The audit of LAION-400M yielded alarming results regarding the prevalence of private data:
- Detection Volume: The pipeline identified 833 unique pregnancy ultrasound images.
- PII Density: From these images, 1,193 instances of private information were extracted, including 387 names and 238 locations.
- Re-identification Risk: Approximately 10% of the images contained multiple types of PII. Most concerningly, 2.1% of the images contained all four categories (Name, Location, Date, and Phone Number), providing enough information for complete re-identification or impersonation of the individuals involved.
Conclusion and Future Directions
The findings suggest that current PII detection is likely an underestimate due to the inherent difficulties of OCR on distorted or handwritten text. This study highlights that pregnancy ultrasounds are likely just one of many sensitive categories—such as X-rays or prescription labels—currently residing in public datasets.
The researchers advocate for a shift in how AI datasets are curated. Rather than "more is better," the community must develop robust tools to filter PII before dataset release and explore "machine unlearning" techniques to remove sensitive data from existing models. Protecting contextual integrity is essential for the ethical advancement of generative AI.
Summary updated