NCorr-FP: A Neighbourhood-based Correlation-preserving Fingerprinting Scheme for Intellectual Property Protection of Structured Data

TL;DR

NCorr-FP tackles IP protection for structured data by embedding recipient-specific fingerprints through a neighborhood-aware, correlation-preserving mechanism that minimizes statistical distortion. By seeding a secret-key driven PRSG, it selects records, attributes, and fingerprint bits to be embedded, sampling new values from density-based regions within correlated neighbourhoods to preserve data fidelity, while enabling blind detection and collusion-resilient tracing via Tardos codes. Empirical results on the Adult Census dataset show fingerprints are nearly imperceptible (Hellinger $<0.023$, KL $<6\times10^{-3}$) and preserve utility (classification accuracy DROP ~1.6%), with high detection confidence and robust performance under single-user and collusion attacks across a range of parameters. The work also provides actionable guidelines for parameter choices, balancing effectiveness, fidelity, utility, and robustness to support real-world deployment on mixed-type data.

Abstract

Ensuring data ownership and traceability of unauthorised redistribution are central to safeguarding intellectual property in shared data environments. Data fingerprinting addresses these challenges by embedding recipient-specific marks into the data, typically via content modifications. We propose NCorr-FP, a Neighbourhood-based Correlation-preserving Fingerprinting system for structured tabular data with the main goal of preserving statistical fidelity. The method uses local record similarity and density estimation to guide the insertion of fingerprint bits. The embedding logic is then reversed to extract the fingerprint from a potentially modified dataset. Extensive experiments confirm its effectiveness, fidelity, utility and robustness. Results show that fingerprints are virtually imperceptible, with minute Hellinger distances and KL divergences, even at high embedding ratios. The system also maintains high data utility for downstream predictive tasks. The method achieves 100\% detection confidence under substantial data deletions and remains robust against adaptive and collusion attacks. Satisfying all these requirements concurrently on mixed-type datasets highlights the strong applicability of NCorr-FP to real-world data settings.

Figures (11)

• Figure 1: Fingerprinting system: (i) fingerprint embedding process implementing fingerprint generator and fingerprint embedding, and (ii) fingerprint detecting process embedding fingerprint detection and a decoding and accusation mechanism.
• Figure 2: NCorr-FP demonstration for continuous attributes: during embedding in a), PDF of target values from a neighbourhood $\mathcal{N}$ for sampling the new value $r'.A_i$ is divided into low- (blue) and high-density areas (orange) using a density percentile $\phi=75\%$. Here we exemplify sampling from a high-density area when $m=1$. The detection process in b) regenerates the PDF of target values from the fingerprinted data. Observe that the distributions in a) and b) are not identical -- this is due to the changes introduced by the fingerprint marks. They are, however, very similar. Therefore, the low- and high-density division of b) closely matches that of the original data in a), and the observed value is correctly classified into the high-density area in this example, therefore, the (correctly) detected mark bit is $m=1$.
• Figure 3: NCorr-FP demonstration for categorical attributes: in a) the frequencies of the target values from the neighbourhood $\mathcal{N}$ are sorted and grouped: the high-frequency group $\mathcal{HD}_\phi$ contains the least amount of most frequent values such its cumulative frequency is at least $1-\phi=25\%$ of the cumulative frequency of all target values (orange). The rest is the low-frequency group $\mathcal{LD}_\phi$ (blue). This example shows the case when $m=1$, i.e. the new (fingerprinted) value is chosen from the high-frequency group. In b) the detection algorithm generates the frequencies from the fingerprinted data which are highly similar to the original (but not necessarily identical). It recognises that the observed value is in $\mathcal{HD}'_\phi$, hence the (correctly) detected mark bit is $m=1$.
• Figure 4: Effectiveness NCorr-FP: Vote Error Rate (VER) on Adult Census dataset for L=128 and N=20.
• Figure 5: Effectiveness of NCorr-FP : Detection confidence DC (solid lines) and false accusation (FAC) for Adult Census data; k=300 and N=20. FAC is shown for two types of fingerprint codes, hash (dotted lines) and Tardos (dashed lines).