Asymmetric Langevin Unlearning Cuts Privacy Costs With Public Data

GK

Gintare Karolina Dziugaite@gkdziugaiteTECH

Our empirical evaluations using variational Rényi divergence and membership inference attacks confirm these theoretical bounds. ALU effectively thwarts privacy attacks while preserving utility under reasonable distribution shifts! See you at poster #3110! 👋