In this work, presented at the ICLR 2026 Conference, the authors share the first systematic study showing that the Synthetic Minority Over-sampling Technique (SMOTE), despite its widespread use, is inherently non-private and can leak sensitive information. They demonstrate that standard evaluation practices fail to detect this leakage. Also introduced are two new attacks, DistinSMOTE and ReconSMOTE, that can perfectly distinguish or even reconstruct real minority records. Theoretical guarantees are provided for these attacks. The findings highlight the need to reconsider the use of SMOTE in privacy-sensitive applications. Read more.
SAS Authors: Georgi Ganev, Reza Nazari, Rees Davison, Amir Dizche, Xinmin Wu, Ralph Abbey, and Jorge Silva
Outside SAS Author: Emiliano De Cristofaro (UC Riverside)
