Organizations today collect and store large amounts of data in various formats and locations, however they are sometimes required to locate all instances of a certain type of data. Data classi cation enables e cient retrieval of information when needed. This work presents a reference implementation for enterprise data classi cation using Semantic Web technologies. We demonstrate automatic discovery and classi cation of Personally Identi able Information (PII) in relational databases, using a classi cation model in RDF/OWL describing the elements to discover and classify. At the end of the process the results are also stored in RDF, enabling simple navigation between the input model and the ndings in di erent databases. Recorded demo link: https://www.research.ibm.com/haifa/info/demos/ piidiscovery_full.htm
Organizations today collect and store large amounts of data in various formats and locations. When an organization is required to meet certain legal or regulatory requirements, for instance to comply with regulations or perform discovery during civil litigation, it needs to nd all the places where the required data is located. Data discovery and classi cation is about nding and marking enterprise data in a way that enables quick and e cient retrieval of the relevant information when needed. Most existing approaches either require re-classi cation of the data each time the organization's policies change, can only be applied to a single data type or format, or only identify prede ned sets of known elds.
In this work we demonstrate the concept of enterprise data classi cation
using Semantic Web technologies described in [
This reference implementation demonstrates the automatic discovery and
classi cation of Personally Identi able Information (PII) stored in relational
databases. The classi cation process starts with creating a model described using
the Resource Description Framework (RDF), containing the entities to discover
and classify as well as additional information that can help the discovery process
(e.g., type and format); this is referred to as a classi cation model. In this demo
we used a model representing PII, but any model that follows the meta-model
described in [
2. Importing database schemas. 3. Discovering and classifying the data according to the classi cation model, using SPARQL and various classi cation algorithms. 4. Representing the results in a way that allows navigation between the classication model and the speci c columns where the information was found.
Our reference implementation is based on the NeOn toolkit [
Using our classi cation tool users can create projects, build and edit models, import existing models (in both cases, the models are validated against the metamodel) and import or create database metadata RDF representations. Users can perform the discovery process on any combination of models and databases. The results, as well as the models and database metadata, can be viewed in both a hierarchical view and a graph view, as depicted in Figure 2. Figure 2 shows part of the discovery results (in this case - all table columns in which a rst name was discovered). For the purpose of this demonstration, we execute our classi cation on two externally available databases: one representing employee records in an organization (taken from the sample database created by the DB2 R software installation) and the other representing medical records of patients (taken from "Avitek Medical Records Development Tutorials" by BEA Systems, Inc. 3). 3 http://download.oracle.com/docs/cd/E13222_01/wls/docs100/medrec_ tutorials/index.html
As noted previously, we use RDF to represent the discovery results, making it possible to navigate from any node in the result back to both the classi cation eld in the model, and to the data eld (column) in the database representation. This easy navigation allows verifying the classi cation results, re ning them (adding or removing triples), and enriching the model so it is more accurate in subsequent runs. 3
In this demonstration we exhibited the main advantages of our approach. By combining di erent discovery techniques and extracting most of the search logic to external les, we created a highly exible and adaptable solution. Using RDF to represent both the ontologies and the results maximizes the modularity and extensibility of the classi cation input and facilitates easy navigation between the results, the models and the data sources. The ontology can thus serve as a centralized point to manage all valuable information in the organization and enables easy location of all related pieces of data in one click. In addition, all of the information created and used by the system (models, metadata RDF representations, results) can be exposed to existing and evolving Semantic Web tools, such as semantic query languages, reasoning engines and rule languages.