Auditing Data Privacy for Machine Learning

Explore the critical issue of data privacy in machine learning through this 18-minute conference talk from USENIX Enigma 2022. Delve into the risks posed by large machine learning models that memorize significant amounts of individual data from their training sets. Learn about inference attacks, particularly membership inference attacks, and their role in measuring information leakage from models. Examine real-world examples from major tech companies and various sensitive datasets to understand the privacy implications. Discover the importance of auditing tools like ML Privacy Meter in assessing and mitigating privacy risks. Gain insights into the differences between privacy and confidentiality, the vulnerabilities of models to inference attacks, and methodologies for quantifying privacy risk. Understand the relevance of these concepts to ML engineers, policymakers, and researchers in developing privacy-conscious machine learning systems.