US 12,093,793 B2
Membership leakage quantification to verify data removal
Abigail Goldsteen, Haifa (IL); and Ron Shmelkin, Haifa (IL)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by INTERNATIONAL BUSINESS MACHINES CORPORATION, Armonk, NY (US)
Filed on Mar. 3, 2021, as Appl. No. 17/191,179.
Prior Publication US 2022/0284341 A1, Sep. 8, 2022
Int. Cl. G06N 20/00 (2019.01)
CPC G06N 20/00 (2019.01) 20 Claims
OG exemplary drawing
 
1. A method for testing data removal from a trained machine learning model trained with a training data set, the method comprising:
training at least one new machine learning model by using an altered data set, the altered data set comprising training data from the training data set, the altered data set being without removal data;
applying a first forgetting mechanism to the trained machine learning model to form at least one first revised machine learning model, the applying of the first forgetting mechanism comprising removing the removal data from the trained machine learning model;
performing a first membership leakage quantification on the at least one first revised machine learning model, wherein the first membership leakage quantification quantifies a first membership leakage of the removal data and uses the at least one new machine learning model for comparison; and
determining a first leakage score from the first membership leakage quantification.