Phantoms and Disclosures: a Causal Framework for Auditing Synthetic Data

The creation of synthetic data using generative AI and Large Language Models poses significant privacy risks, as high-utility synthetic data can inadvertently memorize and disclose private information from the training corpus. Researchers have developed a customizable auditing framework to assess the privacy risks associated with synthetic data generation. This framework aims to identify potential privacy vulnerabilities in synthetic data, enabling the detection of private information memorization. The framework's development is crucial, as the use of synthetic data becomes more widespread, and the need to protect sensitive information grows. The risk of private information disclosure underscores the importance of auditing synthetic data generation processes¹. So what matters to practitioners is that this framework can help mitigate the risk of sensitive information exposure, ensuring the responsible use of synthetic data in various applications.