notesum.ai
Published at December 3DP-2Stage: Adapting Language Models as Differentially Private Tabular Data Generators
cs.LG
cs.CL
cs.CR
D.4.6; G.3; I.2.7
Released Date: December 3, 2024
Authors: Tejumade Afonja1, Hui-Po Wang, Raouf Kerkouche, Mario Fritz
Aff.: 1CISPA Helmholtz Center for Information Security

| Adult | Airline | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 | AUC | ACC | HIST | F1 | AUC | ACC | HIST | ||
| Real data | 69.92 | 91.71 | 84.04 | 99.10 | 90.68 | 96.25 | 91.87 | 99.40 | ||
| CTGAN | 59.56 | 88.50 | 80.23 | 91.21 | 87.23 | 94.72 | 88.93 | 94.41 | ||
| TVAE | 63.22 | 87.51 | 77.74 | 91.51 | 85.85 | 93.06 | 87.26 | 90.32 | ||
| VAE | 53.810 | 86.61 | 80.51 | 73.33 | 79.81 | 91.11 | 80.01 | 76.80 | ||
| GPT-2 | 68.90 | 90.70 | 83.72 | 90.71 | 89.65 | 95.93 | 91.44 | 90.81 | ||