DIPBox: A Multi-Scale Testing Framework for Tracing Dataset Regeneration

The paper proposes DIPBox, a testing framework that traces suspected dataset regeneration through multi-scale similarity tests. The authors' measurement study of public oncology datasets identifies extensive partial dataset copying, and they classify dataset features into three categories—sample-level, set-level, and distribution-level—designing four similarity metrics to identify regeneration. The authors also provide a learning-theoretic analysis formalizing the inherent trade-off between utility and divergence, and infer that there are fundamental limits to evasive regeneration. Experiments on 16 visual and textual foundation datasets, 320 regenerated datasets, and 590 derived models show that DIPBox outperforms prior schemes, and they characterize its robustness and limitations under three adaptive attacks.