What AI says about PrivacyGuard

PrivacyGuard is a library from Facebook Research for privacy analysis of machine learning models, including PyTorch and large language models. It implements various privacy attacks (e.g., membership inference, text inclusion) and provides analysis nodes to interpret results, enabling both quick off-the-shelf assessments and deeper investigations of privacy leakage and memorization. It features an extensible API for adding new analyses and attacks, supports metrics like AUC/ROC and empirical epsilon, and can perform end-to-end privacy attacks including LLM text-generation attacks and probabilistic decoding.