Data traceability and forensic watermarking

Forensic watermarking adds data traceability: it can help identify which delivery a leaked or misused copy may have come from when you share data with a partner, a CRO or a consortium. We create a uniquely marked version for each recipient and record the delivery. If a copy later appears, the watermark is recovered where possible and checked against the register to produce a technical attribution. The strength of that attribution depends on the watermark surviving any transformations of the data and on the delivery register being properly safeguarded, which is why the watermarking combines several techniques so that it does not depend on a single point. It does not prevent copying and cannot guarantee that the source will always be identifiable, but it adds traceability and may deter misuse because deliveries are not identical. Any disciplinary or legal use of the attribution remains a matter for the client and its legal advisers.

Forensic data watermarking is designed to be imperceptible in analytical use and not to change the conclusions a legitimate analysis would be expected to reach. This applies to any delivery that must remain fully useful to the authorised recipient while it carries the watermark. The watermark sits alongside the data without degrading their usefulness. When a numeric value is perturbed, the change is made to the least significant bit (LSB), a magnitude that is negligible for analysis; when a temporal variation is applied, it is designed to preserve the temporal relationships that matter for the planned analysis. The legitimate recipient works with the data without noticing the watermark and without their results changing. That analytical imperceptibility is a design condition, not a lucky side effect. What you receive is a watermark that legitimate use does not perceive and forensic verification can read, not an alteration that compromises the value of the data.

Several complementary techniques, layered, not just one: forensic data watermarking combines different mechanisms so that attribution does not depend on a single point. This applies to tabular data, with the technique chosen according to the type of delivery. On microdata, the watermark can combine decoy rows (honeyrows) that are unique to each recipient and consistent with the rest of the dataset, least significant bit (LSB) perturbation of continuous numeric values, and a variation in temporal precision specific to each recipient. Combining techniques makes the watermark more robust than any of them alone, because it reduces reliance on a single fingerprint. The exact selection is tailored to the structure of each delivery; the same template is not applied to everything. What you receive is a layered watermarking design, unique to each recipient, built on the columns that can carry it without affecting analytical use.

Yes, when the delivery format allows it, with a different mechanism: an aggregate table has no room for decoy rows, so the fingerprint moves to another layer. This applies when what you share are aggregates or summary tables rather than row-level microdata. In that case, the watermarking combines embedded metadata with a canonical fingerprint of the table itself, so that the delivery can still be attributed even though it contains no individual records. The principle is that each type of delivery has its own watermarking mechanism and the same one is not forced onto everything: the system adapts to whether the delivery is microdata or aggregates. This extends traceability to settings where tables are released rather than individual-level data. What you receive is watermarking suited to the nature of the table, not a single technique that only works on microdata.

A documented technical verification that can support an attribution, not a suspicion: the forensic verification procedure checks the dataset examined against the delivery log and shows which delivery it came from and through which fingerprint. This applies when a suspect dataset turns up and its source has to be determined against the record of what was delivered. Verification relies on a digital chain of custody that logs every delivery and keeps a traceable record of the attribution. The procedure is documented and can be re-run, so that it can be reviewed and the technical attribution can be assessed independently. The result shows which delivery the dataset came from and, when one extract went to several recipients, which of them it corresponds to. One limit is worth stating: what you receive is a documented, reproducible technical attribution, not an opinion on its evidential value. Whether that attribution can serve as evidence, and by what route, is for the client and its legal advisers to establish, not for this service.

The data traceability and forensic watermarking service provides a documented, traceable and reproducible technical attribution; it does not establish or vouch for its evidential value in court. This needs stating because "forensic" and "chain of custody" are technical terms of the discipline, not a promise of a legal outcome. What is included: a verification procedure that attributes the source of a delivery in a way that can be re-run, and a digital chain of custody that logs deliveries as an integrity and traceability mechanism. What is not included: claiming that the attribution is legally valid, that it constitutes evidence or that the chain of custody will carry evidential weight. That depends on the applicable legal framework, the jurisdiction and the judgement of the client's legal advisers, and this service does not cover that ground. ADI provides the technical attribution; the decision to pursue misuse, and by what legal route, belongs to the client. Anyone who needs evidential value will find here a documented technical attribution for their legal advisers to assess.

No: no watermarking system is offered as foolproof, and presenting it that way would be dishonest. It is worth understanding this limit before entrusting protection to any single measure. Layered watermarking with several techniques is deliberately more robust than a single technique, because removing or tampering with one fingerprint does not erase the others, but robustness is not invulnerability. There is also a decisive operational condition: how effective the watermarking is in production depends on rigorous management of the keys and the delivery log, which is the client's responsibility. A well-designed system with careless key custody loses its ability to attribute. That is why the service delivers the watermarking design, the verification procedure and custody guidelines, and sets out honestly what each side is responsible for. What we offer is reinforced, well-documented traceability with its conditions of use made explicit, not a promise of infallibility that no serious system can make.

It adds source attribution: data traceability and forensic watermarking tell you which delivery a dataset came from if it leaks, whereas health data anonymisation reduces the risk that people are re-identified. They do not answer the same question: anonymisation asks how to share data while reducing the risk to individuals; traceability asks how to tell which delivery a copy came from when it turns up where it shouldn't. They are sequential and combine naturally, since you can anonymise a delivery and also watermark it to make it attributable, but they are independent services, commissioned separately. One reduces the risk to the people in the dataset; the other strengthens the organisation's ability to attribute the source of a copy. If what you need is to reduce re-identification risk, the service is health data anonymisation; if you need to attribute the source of a leak, it is this one. Applied together, they cover two different risks in the same data release.