Adversarial Synthetic Data: Unmodified Stable Diffusion U-Net Generates Plausible Tabular Data Poisoning Attacks

arXiv·medium signal

Researchers demonstrate that an attacker without resources to train a tabular-specific generator can repurpose an off-the-shelf Stable Diffusion U-Net to generate adversarial synthetic structured data. By reshaping tabular rows into pseudo-images, the architecture's inductive bias toward spatial locality becomes exploitable. This opens a new supply chain attack vector: ground truth drift via publicly available image diffusion models applied to non-image domains.