Understanding and Mitigating Memorization in Diffusion Models for Tabular Data

Zhengyu Fang, Zhimeng Jiang, Huiyuan Chen, Xiao Li, Jing Li · 2024

Read Understanding and Mitigating Memorization in Diffusion Models for Tabular Data by Zhengyu Fang, Zhimeng Jiang, Huiyuan Chen, Xiao Li, Jing Li (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-11044.txt.

Description

Zhengyu Fang and co-authors show that table-generating diffusion models quietly copy their training rows. Their 2024 paper gives the first full study of memorization in this setting. Fake tables are prized for sharing data safely, but copies leak the very privacy they promise.

Memorization grew with more training rounds and shifted with dataset size and width. The authors explain in theory why the copying happens. Their fix swaps chunks of features between rows of the same class during training. A refined version swaps whole clusters of linked features to keep rows realistic. Tests across datasets showed less copying with quality kept.

Catalog record

CREATORZhengyu Fang, Zhimeng Jiang, Huiyuan Chen, Xiao Li, Jing Li
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE131 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-11044

Browse the full library · About the Tower of Babel