Presentation
$\mathbf{C^2PI}$: Crypto-Clear Private Inference
SessionUnleashing Privacy for AI
DescriptionPrivate inference (PI) has addressed the rising concern over data and model privacy in machine learning inference as a service, however, at the cost of high complexity. In this work, we propose to lighten PI by introducing an empirical-based privacy evaluation that leverages inference data privacy attacks (IDPAs) to evaluate data privacy. We then present an enhanced IDPA, DINA, for improved privacy evaluation. Finally, we leverage the findings from DINA and propose $\mathbf{C^2PI}$, which requires only the initial few layers to be performed with PI. Our experiments show that $\mathbf{C^2PI}$ can speed up existing PI frameworks by more than $3$x.
Event Type
Research Manuscript
TimeTuesday, July 11th2:10pm - 2:25pm PDT
Location3004, 3rd Floor
AI
AI/ML Security/Privacy