Back to search

McMaster University

Robust and Fragile Watermarking for Ownership and Integrity Verification of Tabular Data

Abstract

dc:description.abstract

Tabular data are widely shared for machine learning and data analytics. Once released, their owners need to verify ownership of reused data and verify integrity after unauthorized tampering. This thesis develops three watermarking methods for different verification objectives. PKF is a robust watermarking method for ownership verification of tabular data used to train machine learning models. Without relying on primary keys, which adversaries can modify or remove without degrading utility, PKF maps numerical attributes to a discrete-time signal, embeds a sinusoidal watermark, and detects its spectral evidence. Its analysis bounds the attack-induced perturbation under which detection succeeds with high probability. Experiments on two datasets show superior robustness over six baselines under six watermark-removal attacks. RaMark is a robust watermarking method for ownership verification of tabular data generated by generative models. It introduces radioactivity, the property that a watermark remains detectable under a retraining attack, in which a generative model is trained on watermarked data and used to generate new tuples. RaMark embeds a sinusoidal dependency into the data distribution and detects it through spectral analysis. The analysis shows that when regenerated and watermarked data distributions remain close in Wasserstein-1 distance, their spectral powers remain close with high probability. Experiments on two datasets show stronger robustness than seven baselines under retraining attacks. LoMark is a fragile watermarking method for integrity verification of relational databases under sparse tampering. It uses primary key set reconciliation to localize inserted and deleted tuples and group testing to localize value-modified tuples. Theoretical guarantees characterize when localization succeeds with high probability and guide parameter selection under a verification metadata budget. Experiments on four datasets against eight baselines show accurate tuple-level localization using verification metadata amounting to 1% of the database size. Overall, this thesis provides watermarking methods for ownership and integrity verification of released tabular data.

Degree

thesis:*
Department dc:contributor.department
Computing and Software
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Xin Che
Advisor dc:contributor.advisor
  • Lingyang, Chu

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
McMaster University
Base URL
macsphere.mcmaster.ca/server/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Xin Che. Robust and Fragile Watermarking for Ownership and Integrity Verification of Tabular Data. 2026. https://hdl.handle.net/11375/33888