Systematic Evaluation of Dataset Watermarking for Intellectual Protection

Summary

This study reassesses nine dataset watermarking methods from 2025 under a unified experimental framework, splitting them into prediction-consistent and prediction-inconsistent groups. Most methods cause only minor overall accuracy loss (under 1%), but clean-label methods can reduce target-class accuracy by as much as 3.6%. Under a data mixture scenario, one optimization-based method and Style Transformation fail completely, and all methods except the domain watermark show utility drop, weakened verification, or poor robustness.