Research map / Image generation / Fast diffusion sampling

Understanding and Mitigating Memorization in Diffusion Models for Tabular Data

ICML 2025 · Image generation › Fast diffusion sampling · approach: Diffusion models as perception tools · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Investigates memorization in tabular diffusion models (TabSyn, TabDDPM, STaSy), explains it theoretically via the optimal score, and proposes TabCutMix and TabCutMixPlus cross-sample feature-swap augmentations that reduce memorization while keeping synthetic data quality.

Problem

Memorization, where tabular diffusion models replicate near-identical training records, is largely unexplored for tabular data, and it raises privacy and generalization concerns.

Approach

Adapt CutMix to tabular data by swapping random feature segments between same-class training pairs (TabCutMix), and cluster correlated features so that swaps keep correlated features together (TabCutMixPlus) to reduce out-of-distribution generation.

Key results

• TabCutMix reduces TabSyn memorization ratio on Default from 20.11% to 16.86% (16.16% relative) and on Shoppers from 27.68% to 25.38% (aside from TCMP results).
• TabCutMixPlus keeps higher data quality than TabCutMix in several metrics, e.g., Default TabSyn alpha-precision 97.61% vs 96.16% for TCM.
• Memorization ratio increases with training epochs and is larger for smaller training sets, with Magic reaching about 80% memorization for TabSyn.

Measured effects reported in the paper

Surprising findings

• Under the paper's own metrics, TabCutMix reduces memorization more than TabCutMixPlus, though TabCutMixPlus gives better data quality.
• Mixup and SMOTE sometimes reduce memorization but at large cost to C2ST and shape/trend scores (e.g., SMOTE trend score 32.70% on Default TabSyn).

Limitations

• Theoretical result relies on idealized assumptions (perfect score network, infinitesimal SDE step) and is only partly applicable to the VAE-based TabSyn.
• Memorization is measured by a distance-ratio threshold heuristic adapted from image generation rather than a verbatim criterion.
• Limited to four classification datasets in the main analysis and no public code release statement beyond a GitHub link.

Experimental setup

Datasets: Adult Default Shoppers Magic

Benchmarks: Adult Default Shoppers Magic

Metrics: memorization ratio Mem-AUC MLE AUC alpha-precision beta-recall shape score trend score C2ST DCR

Baselines: Mixup SMOTE STaSy TabDDPM TabSyn CTGAN TVAE

Builds on: CutMix TabDDPM TabSyn STaSy

Closest papers by meaning

• TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.90)
• How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.88)
• Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.88)
• On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.88)
• Continuous Diffusion for Mixed-Type Tabular Data (ICLR 2025 · similarity 0.87)
• On the Memorization of Consistency Distillation for Diffusion Models (NeurIPS 2026 · similarity 0.87)