arXiv:2510.24233v1 Announce Type: new 
Abstract: Deep generative models are often trained on sensitive data, such as genetic sequences, health data, or more broadly, any copyrighted, licensed or protected content. This raises critical concerns around privacy-preserving synthetic data, and more specifically around privacy leakage, an issue closely tied to overfitting. Existing methods almost exclusively rely on global criteria to estimate the risk of privacy failure associated to a model, offering only quantitative non interpretable insights. The absence of rigorous evaluation methods for data privacy at the sample-level may hinder the practical deployment of synthetic data in real-world applications. Using extreme value statistics on nearest-neighbor distances, we propose PRIVET, a generic sample-based, modality-agnostic algorithm that assigns an individual privacy leak score to each synthetic sample. We empirically demonstrate that PRIVET reliably detects instances of memorization and privacy leakage across diverse data modalities, including settings with very high dimensionality, limited sample sizes such as genetic data and even under underfitting regimes. We compare our method to existing approaches under controlled settings and show its advantage in providing both dataset level and sample level assessments through qualitative and quantitative outputs. Additionally, our analysis reveals limitations in existing computer vision embeddings to yield perceptually meaningful distances when identifying near-duplicate samples.

تقدم دراسة جديدة مقياس PRIVET، وهو مقياس للخصوصية يعتمد على نظرية القيم المتطرفة، حيث يتناول القضية الحرجة لتسرب الخصوصية في النماذج التوليدية العميقة المدربة على بيانات حساسة مثل سجلات الصحة والتسلسلات الجينية. هذه الابتكار مهم لأنه يسعى إلى تعزيز أمان البيانات الاصطناعية، التي تُستخدم بشكل متزايد في مجالات متنوعة. من خلال التركيز على تقنيات الحفاظ على الخصوصية، يمكن أن يساعد PRIVET في التخفيف من المخاطر المرتبطة بالتكيف المفرط وضمان بقاء المعلومات الحساسة محمية، مما يجعله تطورًا حيويًا في مجال خصوصية البيانات.

Un nuevo estudio presenta PRIVET, una métrica de privacidad basada en la teoría de valores extremos, que aborda la crítica cuestión de la fuga de privacidad en modelos generativos profundos entrenados con datos sensibles como registros de salud y secuencias genéticas. Esta innovación es significativa ya que busca mejorar la seguridad de los datos sintéticos, que se utilizan cada vez más en diversos campos. Al centrarse en técnicas de preservación de la privacidad, PRIVET podría ayudar a mitigar los riesgos asociados con el sobreajuste y garantizar que la información sensible permanezca protegida, lo que lo convierte en un desarrollo vital en el ámbito de la privacidad de datos.

Une nouvelle étude présente PRIVET, une métrique de confidentialité basée sur la théorie des valeurs extrêmes, abordant la question cruciale de la fuite de données dans les modèles génératifs profonds formés sur des données sensibles telles que les dossiers de santé et les séquences génétiques. Cette innovation est significative car elle vise à améliorer la sécurité des données synthétiques, de plus en plus utilisées dans divers domaines. En se concentrant sur des techniques de préservation de la vie privée, PRIVET pourrait aider à atténuer les risques associés au surajustement et garantir que les informations sensibles restent protégées, ce qui en fait un développement essentiel dans le domaine de la confidentialité des données.

A new study introduces PRIVET, a privacy metric based on extreme value theory, addressing the critical issue of privacy leakage in deep generative models trained on sensitive data like health records and genetic sequences. This innovation is significant as it aims to enhance the safety of synthetic data, which is increasingly used in various fields. By focusing on privacy-preserving techniques, PRIVET could help mitigate risks associated with overfitting and ensure that sensitive information remains protected, making it a vital development in the realm of data privacy.

PRIVET: Privacy Metric Based on Extreme Value Theory

One More Thing in AI – Your Shortcut to AI Mastery

PRIVET: Privacy Metric Based on Extreme Value Theory

Was this article worth reading? Share it

One More Thing in AI

GPTHuman

Omnifact

Ask On Data

GPTHumanizer

Adaptive Privacy Policy Generator

Ready to build your own newsroom?