Beyond One-Size-Fits-All: Neural Networks for Differentially Private Tabular Data Synthesis

In differentially private (DP) tabular data synthesis, the consensus is that statistical models are better than neural network (NN)-based methods. However, we argue that this conclusion does not fully capture algorithms' performance across different data regimes, particularly when correlations are densely distributed. In such complex scenarios, intricate dependencies can overwhelm statistical models, and NNs may be more competitive due to their capacity to fit complex distributions by learning directly from samples. Therefore, to fully evaluate algorithms across diverse data characteristics, we propose MargNet, which integrates adaptively selected marginals into NN training to better realize their generative capacity, and conduct evaluations on various datasets. On sparsely correlated datasets, although the prior state-of-the-art statistical baseline AIM continues to achieve high utility, MargNet achieves comparable performance with a significant speedup. On densely correlated datasets, MargNet achieves the best synthesis utility in most settings. These findings suggest that NNs remain an important approach to DP tabular data synthesis and that algorithm selection should account for dataset characteristics. We released our source code on GitHub. 1

• Security and privacy → Privacy-preserving protocols.