Dataset Usage Inference without Shadow Models or Held-out Data
Wojciech ŁapaczWarsaw University of TechnologyStanisław Pawlak Warsaw University of TechnologyJan Dubiński Warsaw University of TechnologyNASK National Research InstituteFranziska BoenishCISPA Helmholtz Center for Information SecurityAdam DziedzicCISPA Helmholtz Center for Information Security · 2026
Read Dataset Usage Inference without Shadow Models or Held-out Data by Wojciech ŁapaczWarsaw University of TechnologyStanisław Pawlak Warsaw University of TechnologyJan Dubiński Warsaw University of TechnologyNASK National Research InstituteFranziska BoenishCISPA Helmholtz Center for Information SecurityAdam DziedzicCISPA Helmholtz Center for Information Security (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-26257.txt.
Description
Wojciech Łapacz and colleagues answer a question haunting generative AI: how much of my dataset trained your model? Dataset usage inference estimates that fraction for a candidate dataset. Older methods demand shadow models, costly lookalikes trained to mimic the target. They also demand confirmed non-member data from the same distribution. Both demands collapse in real ownership disputes. Their framework drops both requirements at once.
Synthetic stand-ins replace the held-out data. The method generates its own non-member samples, then extracts varied membership signals from the model. Usage becomes a mixture proportion problem: what share of the candidate set behaves like training data? Experiments on large image generative models show the estimates track the truth reliably. Data owners gain a practical audit tool that works on today's giant models. Copyright and consent disputes, like the lawsuits now moving through courts, are the intended courtroom for such evidence.
Catalog record
| CREATOR | Wojciech ŁapaczWarsaw University of TechnologyStanisław Pawlak Warsaw University of TechnologyJan Dubiński Warsaw University of TechnologyNASK National Research InstituteFranziska BoenishCISPA Helmholtz Center for Information SecurityAdam DziedzicCISPA Helmholtz Center for Information Security |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.26257