missing. Most privacy-preserving schema matching methods are based on the assumption that direct access to plaintext data is not possible. Consequently, only metadata, such as attribute names and their associated descriptions, can be utilized. However, in many cases, even these metadata are not fully available, with attribute descriptions often Our claim is that privacy-preserving instance-based schema matching can still be achieved by leveraging anonymized data. Specifically, we propose to evaluate how -anonymity influences the performance of instancebased schema matching. To this end, we investigate the efect of -anonymity on the accuracy and eficiency of instance-based schema matching methods.
Data integration across organizations ofers significant advantages, such as enhancing data quality
and uncovering new insights that individual databases alone cannot provide [
As stated in [
In this paper, we propose a Privacy-Preserving Schema Matching technique that builds upon wellestablished results and existing frameworks from the literature in both privacy protection and schema matching, integrating them into a unified approach.
Regarding privacy protection, anonymization aims to prevent the re-identification of individuals by ensuring that data records can no longer be uniquely traced back to a specific person. Among various privacy paradigms, k-anonymity is the most widely adopted. It works by transforming quasi-identifiers (QIDs)—such as age, gender, or location—so that each record becomes indistinguishable from at least −1 others in the dataset. However, even this relatively simple privacy model can introduce data distortion, leading to information loss and a potential negative impact on automated analysis—including techniques such as schema matching. The extent of this distortion directly afects data usability, making it essential
CEUR Workshop
ISSN1613-0073
to minimize information loss. For this reason, inspired by the findings presented in [
With respect to schema matching, existing techniques can be grouped according to the type of information they leverage. Specifically, schema-level approaches rely solely on schema-level metadata—such as attribute names, data types, and contextual information—to identify potential matches. In contrast, instance-level methods focus on the actual data values, using strategies such as comparing attribute value distributions or applying syntactic similarity metrics.
Instance-level schema matching often violates privacy constraints because it requires querying actual
data on both sides. To address this, SMAT [
However, this method presents two main limitations. First, it requires the disclosure of local table and attribute names, along with their descriptions—an assumption that may not be acceptable in privacypreserving scenarios. Second, in many real-world cases, such metadata are incomplete or unavailable, with attribute descriptions often missing. For these reasons, the privacy-preserving schema matching technique we propose is purely instance-based, meaning it does not rely on table or attribute names or descriptions, which can therefore be treated as meaningless symbols. Instead, it operates directly on anonymized data.
The main contribution of this paper is a Privacy-Preserving Schema Matching technique that leverages well-established results and builds on existing frameworks in the literature across both privacy protection and schema matching, combining them into a unified approach. More precisely, we present preliminary results obtained by applying classical instance-based schema matching methods in conjunction with well-known -anonymity techniques. To assess the impact of anonymization on schema matching performance, we go beyond a standard evaluation against a known gold standard by also comparing the results obtained from -anonymized data to those derived from cleartext datasets. This comparison allows us to quantify the efect of -anonymity on result quality, even in the absence of a reference alignment.
The paper is organized as follows.
Section 2 introduces the foundational components of the proposed Privacy-Preserving Schema Matching approach. In particular, we describe the classical instance-based schema matching methods that serve as the core matching algorithms, and the well-known -anonymity techniques employed to protect data privacy. These elements, drawn from well-established literature in both schema matching and privacy preservation, are combined into a unified framework designed to support schema alignment tasks while minimizing disclosure risks. In the Experimental Setup section 3, we present both the Dataset Preparation, i.e., the datasets used in our experiments, and the Experimental Workflow employed. Section 4 - Experimental Results presents some results. Section 5 is devoted to the related works. Section 6 contains conclusions, and directions for future work.
This section introduces the foundational components of the proposed Privacy-Preserving Schema
Matching approach. We describe some classical instance-based schema matching methods implemented in
the Valentine Framework [
In this section, we briefly describe the Valentine framework [
Valentine is an extensible open-source experiment suite designed to execute and organize large-scale automated schema matching experiments on tabular data. It allows the application of multiple schema matching methods to pairs of denormalized tables, leveraging available metadata such as table and attribute names, data types, and the actual data values.
An important feature of Valentine is that it produces a ranked list of matching attribute pairs, ordered by the confidence score assigned by each method. This is particularly suitable for dataset discovery scenarios, where ranked results allow users and systems to explore candidate matches more eficiently and assess match quality based on position in the ranking.
To evaluate the efectiveness of the methods, Valentine introduces the Recall@ground truth metric, defined as follows:
This metric reflects how many of the correct matches appear in the top- results returned by a method, thus providing an intuitive measure of how helpful the ranked list is for a human evaluator or downstream process that considers only a limited number of top suggestions.
Below is a brief description of the instance-based matchers by Valentine utilized in our work. Distribution-based Matcher.
Distribution-based Matching [
Jaccard Distance Matcher.
This is an instance-based matcher that uses Jaccard to calculate all pairwise column similarities; two values are considered identical if their Levenshtein distance (or other similarity measure) is less than a certain threshold.
Access to data is fundamental for open science, supporting transparency, reproducibility, and research progress. However, privacy regulations like the GDPR [11] restrict the publication and sharing of datasets containing sensitive information.
A key challenge is the risk of re-identification, where individuals can be recognized even if direct identifiers (e.g., names) are removed. This risk increases when datasets include quasi-identifiers , such as age, postal code, and occupation, which combined can uniquely identify individuals. To reduce this risk, k-anonymity requires that each combination of quasi-identifier values appears in at least records, making each individual indistinguishable from −1 others. This reduces re-identification chances but causes some information loss.
K-anonymity has known limitations, including vulnerability to homogeneity and background knowledge attacks, which happen when all records share the same quasi-identifier values.
Optimal Lattice Anonymization.
Optimal Lattice Anonymization (OLA) [12] is an algorithm for achieving -anonymity by generalizing and suppressing data to protect sensitive information. It ensures each record is indistinguishable from at least − 1 others, reducing re-identification risk. OLA searches a lattice of possible generalizations, where each node represents a diferent way of generalizing attributes. Using predictive tagging, it eficiently prunes the search by skipping nodes that are guaranteed to satisfy or violate -anonymity, cutting down computations. After identifying all -anonymous nodes, OLA keeps only those with minimal generalization and selects the best one based on information loss measures like the Discernibility Metric , choosing the solution with the least loss of data utility. Mondrian. The Mondrian algorithm [13] is a greedy top-down recursive partitioning method for achieving -anonymity. It partitions the multidimensional quasi-identifier space into regions containing at least indistinguishable records.
At each step, the attribute with the largest range is selected and the domain is split along the median, proceeding recursively until the -anonymity constraint is satisfied or no further splits are possible. Mondrian supports both global generalization, which uniformizes the treatment of instances, and local generalization, which preserves specific details in some partitions, balancing privacy and utility. This approach, characterized by computational eficiency and simplicity, is widely adopted in practical data anonymization applications.
The goal of our experimental study is to assess how -anonymity impacts the performance of instancebased schema matching algorithms. The evaluation metrics used are those defined in section 2.1.
The first type of evaluation adopted consists of classical comparison against a known gold standard. This mode, which we refer to as Evaluation against the Gold Standard, is applied to both the results obtained on the cleartext data and those derived from the k-anonymized data.
The second type of evaluation, which we will call Evaluation against Cleartext Results, compares the
results obtained on k-anonymized data to those obtained on the original cleartext datasets, treating
the latter as a gold standard proxy. This type of evaluation is strongly inspired by the methodology
proposed in [
The technique we adopted to construct the datasets used in our experiments is inspired by Valentine’s
[
The dataset preparation procedure is the following.
(1) Selection of the starting dataset. The starting point is a real dataset commonly used in studies on k-anonymization, namely the Adult Dataset, which contains 45,222 records derived from the 1994 United States Census. We consider the following attributes: sex, age, race, marital-status, education, native-country, workclass, and salary-class. This dataset is particularly suitable because it is already integrated into existing frameworks for the evaluation of k-anonymization, and it has known and available generalization hierarchies useful for applying anonymization techniques.
(2) Generation of datasets to be compared. Starting from the original dataset, to make the task more realistic, dummy columns were created for some attributes that are derived from the original ones but with diferent distributions. For example, two new columns, native-country_1 and native-country_2, were added for the native-country attribute; these columns can be interpreted as representing two diferent country-related contexts: the country of current residence and the country of work, with values initially extracted from the original column but modified to have diferent distributions. Similar operations have been performed on other columns, creating new variants with slightly perturbed distributions. Then we build pairs of datasets ( , ) through a controlled horizontal overlap (i.e., a certain percentage of records shared between the two datasets) and a controlled vertical overlap (i.e., a certain percentage of shared attributes).
(3) Perturbation of the datasets. To simulate realistic noisy data conditions, we apply a phase of controlled perturbation to the generated datasets. In particular, we use the GeCo tool [14] to introduce errors into the records, allowing a variable number of errors per record. Data corruption is performed based on the attribute type: for textual attributes, random typos are introduced based on keyboard key proximity; for numeric attributes, values are randomly modified following the statistical distribution of the original values. This phase allows simulating syntactic or semantic noise commonly present in real data, and evaluating the robustness of schema matching algorithms in the presence of such perturbations.
The experimental workflow consists of the following steps: 1. Matching on cleartext data: For each dataset pair we apply a schema matching algorithm to the original data and compute evaluation metrics using a known gold standard. We evaluate the following two instance-based schema matching methods, both available within the Valentine framework and described in section 2: DistributionBased and JaccardDistanceMatcher. 2. Data anonymization: We apply well-known -anonymity algorithms (such as Mondrian or
OLA) to generate multiple anonymized versions of each dataset by varying the value of . 3. Matching on anonymized data: Each instance-based schema matcher is applied to the anonymized datasets. 4. Performance comparison: We perform both Evaluation against the Gold Standard and Evaluation against Cleartext Results, in order to assess how -anonymization afects the efectiveness of schema matching methods under diferent evaluation perspectives.
In this section we present and discuss some preliminary results.
Cleartext Data - Evaluation against the Gold Standard To evaluate the performance of matchers (JaccardMatcher and DistributionMatcher, see section 2.1) on the plaintext data, we use the classic comparison of results with a Gold Standard and calculate the Recall@GroundTruth (see Table 1).
Pair ( 1, 2) ( 3, 4)
JaccardMatcher 0.57 0.67
DistributionMatcher 0.71 0.92
Best Matcher Distribution Distribution
Overall, the DistributionMatcher proves to be the most efective in all two cases, demonstrating better quality in matching tasks compared to JaccardMatcher on cleartext data. This is to be expected, since dummy columns were created with slightly modified distributions: while Jaccard is based on overlap of values, Distribution captures similarities at the distribution level. Anonymized Data - Evaluation against the Gold Standard To assess which k-anonymizer between OLA and Mondrian allows for better quality schema matching tasks, we compared the values of Recall@GroundTruth obtained against the Gold Standard, for each dataset pair combination and for each matcher used (see Table 2).
Matcher JaccardMatcher JaccardMatcher DistributionMatcher DistributionMatcher
The analysis suggests that the quality of schema matching tasks with respect to the Gold Standard is better when done on anonymized data with OLA, for both the JaccardMatcher and DistributionMatcher. Regarding the efect of the matcher used, OLA improves more with the DistributionMatcher than with the JaccardMatcher.
Figure 1 summarizes the results for this evaluation.
Evaluation against Cleartext Results In this analysis, the JaccardMatcher and DistributionMatcher were considered as two semantic similarity tasks between pairs of datasets: the objective is to test which of the two k-anonymizers, OLA or Mondrian, is more efective in preserving the original similarity as measured by these tasks.
For each matcher JaccardMatcher and DistributionMatcher, for each pair of datasets, for each kanonymizer, we compared the matches obtained between the k-anonymized datasets with those obtained between the plain-text datasets, calculating the Recall@GroundTruth (see Table 3).
We can see, the trend is very similar to the previous one: also here, OLA is always better than Mondrian regardless of the matcher adopted and of the dataset pairs used. This happens because OLA checks all possible generalizations and then selects the one with the best utility.
Figure 2 summarizes the results for this evaluation.
Finally, to evaluate the impact of k-Anonymity on matching quality, that is, to analyze how the degree of anonymization (parameter in -anonymity) influences instance-based schema matching quality, we compared the results of matching on -anonymized data against the results obtained on cleartext data.
Mondrian 0.43 0.50 0.43 0.58
The paper [15] presents a privacy-preserving approach to schema and data matching, where sensitive information is protected through embedding techniques that enable matching operations without revealing original values. As discussed in the paper, the evaluation of the performance of matching techniques on anonymised data is carried out by comparing precision and recall metrics obtained in the embedded space (i.e., after anonymisation) with those obtained in the original space (i.e., without anonymisation).
In [
PRISMA (Privacy-Preserving Schema Matcher) [17] is based on a privacy-preserving schema matching method that avoids access to raw data and relies only on metadata, such as functional dependencies and frequency distributions; it represents schemas as graphs and generates embeddings to compare attributes and identify one-to-one correspondences. Like our method, PRISMA also performs privacypreserving schema matching without accessing the cleartext data. PRISMA, however, employs specific tools developed specifically for this purpose, while our method relies on classical schema matching techniques combined with well-established k-anonymity methods.
A recent trend in schema matching research explores the use of Large Language Models (LLMs) to improve both matching accuracy and scalability. For instance, the Magneto framework [18] proposes a two-phase approach that combines small and large language models to eficiently retrieve and rerank candidate matches. Its architecture relies heavily on both schema-level metadata (e.g., column names) and instance-level data (e.g., sampled values). In [19], the authors propose leveraging LLMs—such as GPT-3.5 and GPT-4—to generate semantic correspondences between attributes of heterogeneous schemas in the healthcare domain. The proposed method relies exclusively on textual metadata (e.g., attribute names and descriptions), deliberately avoiding access to instance-level data in order to reduce exposure to sensitive information.
The use of external LLM services, which entails transmitting data to third-party providers (e.g., OpenAI APIs), violates the requirements of a strictly privacy-preserving scenario, where no sensitive information—neither schema-related nor data-related—should be exposed outside the local environment. While this issue can be mitigated by deploying LLMs in a local environment, such a solution poses significant challenges in terms of computational resources, model maintenance, and the need for technical expertise. Moreover, locally deployed models may not always match the performance or upto-dateness of cloud-based LLMs, potentially limiting their efectiveness in complex schema matching tasks.
In this paper, we presented a preliminary study on privacy-preserving instance-based schema matching using -anonymized data. Our results show that it is possible to achieve efective schema alignment while protecting sensitive information by operating exclusively on anonymized datasets, without relying on schema metadata that may be incomplete or unavailable.
We also demonstrated how comparing matching results on anonymized versus cleartext data provides valuable insights into the trade-of between privacy and data utility in schema matching scenarios.
As future work, we plan to extend our evaluation to real-world clinical datasets [
Another possible future work involves the study of the trade-of between data utility and privacy protection. By data utility we mean how useful a dataset continues to be after applying protection techniques, such as k-anonymization. In general, the greater the protection applied, the greater is also the loss of useful information. Therefore, it is important to understand how to find a good trade-of between privacy and usefulness. In this work, this balance was evaluated indirectly by considering how the performance of schema matching algorithms changes when moving from the original data to anonymized data. Conduct a systematic assessment of the privacy-utility tradeof, measuring both privacy gains (how much the risk of re-identification is reduced) and utility losses, not only in terms of the accuracy of the matching scheme but also with information loss metrics, so as to have more guidance on appropriate k values in practice.
During the preparation of this work, the authors used ChatGPT to check grammar and spelling. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.