Data classification: the first step of data loss prevention

A DLP rule only works once someone has decided how sensitive each piece of data is, and the file or email carries that decision with it.

Topic
IT security
Author
ARLITECH expert team
Published

A data loss prevention (DLP) system asks the same question for every outgoing email, upload, and print job: may this data go where the user is sending it? It can only answer if someone has already decided how sensitive each piece of data is and who may do what with it. That decision is called data classification.

Why DLP does not work without classification

DLP rules can rely on two things: a pattern found in the content, such as a payment card number, and a label attached to the data. Without classification, only the first remains. Pattern matching works well for well-formatted data, but a bid price calculation, a production recipe, or a management presentation contains no character string that would tell a machine it is confidential.

The bigger problem is the missing yardstick. If nothing in writing says what counts as confidential, the administrator writes the rule according to their own judgment. A strict rule also catches many harmless emails, so users look for workarounds, such as their personal email address. A loose rule quietly lets through everything it does not recognize. In neither case can anyone tell afterwards whether the system is doing its job, because there is nothing to measure it against.

Classification provides that yardstick. It states where data of a given category may go, who may see it, and with what protection, and DLP enforces that decision. Annex A of ISO/IEC 27001:2022 also lists classification of information (5.12) and labeling of information (5.13) as separate controls.

How many levels you need: from public to strictly confidential

In enterprise practice, schemes with three to five levels have become the norm. A four-level scheme might look like this:

LevelWhat it meansExamplesHandling
PublicIntended for publication; anyone may see it.Website, product sheet, press releaseNo restrictions.
InternalCan move freely within the organization but is not meant for outsiders.Internal policies, phone directory, internal newsletterTo external recipients only with a justification.
ConfidentialOnly the people involved may see it; disclosure causes harm.Contracts, bids, client lists, payrollNamed recipients, encrypted transfer.
Strictly confidentialA few named people have access; disclosure causes serious harm.Acquisition negotiation documents, source code, production recipesExternal sharing prohibited, encryption, every access logged.

The test for the number of levels is whether an employee looking at a document can decide within a few seconds where it belongs. With seven or eight levels that no longer works: when unsure, users pick the lowest or the highest, and the classification loses its meaning. Two levels are too few, because then the payroll and the internal phone directory end up in the same category.

The levels have to differ in their handling rules. If internal and confidential data can be emailed, printed, and copied to a USB stick in exactly the same way, the two are really one level. For each level, write a short handling table that says what to do when storing, sharing, printing, and disposing of the data.

The classification decision belongs to the data owner, meaning the person responsible for the process in which the data is created: HR for the payroll, finance for the draft financial statements. IT carries out that decision. If IT writes the scheme on its own, it ends up built on technical considerations and leaves out how much business harm is at stake.

What counts as sensitive data

Three questions decide how sensitive a piece of data is: whether a law or a contract protects it, how much harm its disclosure would cause, and whether it gives access to other data.

Personal data

The European Union's General Data Protection Regulation (GDPR) treats as personal data any information relating to an identified or identifiable natural person. Within that, Article 9 lists special categories, including health and genetic data, biometric data used to uniquely identify natural persons, data concerning a person's sex life or sexual orientation, and data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership. These go to the top level of the scheme. Numbers used to identify individuals also belong at a high level: the identity card number, the tax identification number, and the TAJ number (the Hungarian social security number).

Financial data

Payment card data falls under the PCI DSS standard of the PCI Security Standards Council, which places separate requirements on organizations that accept card payments. Bank account numbers, salary data, and financial statements not yet published also belong at the confidential level.

Trade secrets

Act LIV of 2018 on the protection of trade secrets treats information as a trade secret only if, among other conditions, the holder of the secret has taken the steps that can reasonably be expected in the circumstances to keep it secret. A classification and the handling rule that goes with it are such a step, and in a dispute they also help show that the company really did treat the data as secret. This category includes source code, technical documentation, production parameters, and bid prices.

Credentials

Passwords, API keys, private keys of certificates, VPN configurations. On their own they have little business value, but whoever obtains them gets access to the rest of the data. They often end up by accident in tickets, chat messages, and code repositories, so pattern matching has to look for them too.

Aggregation matters as well. An employee's name on its own may be internal data, but a thousand names with tax identification numbers and home addresses in one spreadsheet sits at one of the highest levels. The scheme should state that a collection of data is classified by its most sensitive element and, for large volumes, one level higher.

Manual labeling and automatic detection

The classification has to be attached to the file or email, otherwise DLP cannot see it. There are two ways to do this, and the best results come from using both together.

Manual labeling

The user selects the label when saving the document or sending the email. The label is written into the file's metadata, and many systems also add a header, footer, or watermark to the document, so the classification shows on the printed page too. The person who wrote the content knows it best, so a manual label is usually accurate. Its weak point is haste: if a choice is required at every save, after a while the user clicks without thinking.

A default label helps here. Anything created without a label automatically becomes internal, and the user only has to mark the exceptions. Anyone who moves a label to a lower level has to give a reason, and that goes into the log.

Automatic detection

Here the system infers the classification from the content. The simplest tool is the regular expression, which searches for a character string in a given format. On its own it produces many false positives, because a nine-digit number can be many things. Accuracy improves when the pattern also comes with a checksum: the Luhn algorithm for payment card numbers, the mod 97 check for international bank account numbers (IBAN), and the built-in check digit for the tax identification number and the TAJ number filter out chance matches. It narrows further if the rule also expects a keyword near the pattern, such as “TAJ” or “tax ID”.

Data with no fixed pattern needs other methods. Fingerprinting builds a digest of a sample document, such as a contract template or an internal form, and also recognizes partly modified copies and excerpts. Exact data match looks for the actual values from a client or employee table, so it flags only the names that appear in that table. Machine learning classifiers recognize document types, such as CVs or source code; test them on your own documents before building rules on them. In scanned documents and screenshots the text exists only as an image, and these can only be inspected with optical character recognition (OCR).

The two methods also check each other. If automatic detection finds personal data in a document labeled public, the system can warn the user or alert the data owner.

Channels: email, cloud storage, printers, USB

Data can leave the organization by several routes, and each is visible to a different control point. The classification is the same on every channel, but the rules have to be written channel by channel.

Email

DLP running on the mail server or the email gateway inspects the outgoing message and its attachments. This channel offers the most options: the message can be stopped, quarantined, or sent on with enforced encryption, or the sender gets a warning and can still send it with a justification. The rule has to distinguish internal from external recipients. Inspection cannot see inside password-protected attachments, so you need a separate decision on whether they may pass.

Cloud storage and web uploads

In corporate cloud storage, DLP checks stored files and shares, and raises an alert when someone creates an external link, or a link anyone can open, to a confidential document. For uploads to non-corporate services, whether personal storage or a web-based AI tool, visibility comes from the endpoint agent or from the proxy and firewall that inspect web traffic. This requires decrypting HTTPS traffic (SSL/TLS inspection), which needs its own planning because of certificate management and privacy exceptions. What data can end up in AI tools is covered in the article on shadow AI.

Printers

A printed page is outside digital control. The endpoint agent can allow, block, or log a print job based on the label and the content. At the printer, authenticated printing, where the job only prints after the user presents a card or enters a PIN, keeps a confidential page from sitting in the tray for hours.

Removable media

USB sticks and external hard drives can be controlled on the endpoint. They can be blocked, allowed read-only, restricted to encrypted devices issued by the company, or the policy can block copying only for files labeled confidential. Where a complete ban is not workable, a list of approved devices and logging of copy operations strike an acceptable balance.

Monitor mode, false positives, and gradual enforcement

A new rule starts by only monitoring: it logs what it would stop but blocks nothing. This phase shows what it actually finds. Wait at least one full month, so that processes that recur monthly, such as payroll traffic, also show up in the log.

False positives

Reviewing the matches typically shows that some of the alerts are real: those are the reason the rule exists. Others are wrong, such as a part number that looks like a payment card number, or a developer file full of test data. For these, the pattern has to be narrowed or refined with a keyword. Finally, there are legitimate business processes that really do send sensitive data out, such as the monthly payroll going to an external payroll provider. These first have to be moved to a secure channel, and only then do they get an exception.

Exceptions need the same discipline as firewall rules: each should have an owner, a justification, and an expiry date. A narrow exception for one sender, one recipient domain, or one folder is manageable. If a whole department gets an exception, the rule quickly becomes an empty shell.

Enforcement steps

  1. Monitoring: the rule only logs, and the matches are reviewed together with the data owners.
  2. Warning: the user gets a message saying they are about to send sensitive data and can continue with a justification. The justifications reveal where a secure channel is missing.
  3. Blocking at the top level: sending strictly confidential data out is prohibited, while the other levels still only warn.
  4. Expansion: blocking gradually extends to the confidential level and to more channels.

Before each step, tell users what is changing and whom to contact if a legitimate email gets held up. If a bid gets stuck right before a submission deadline and there is no quick way to release it, it can easily end with the whole rule being switched off.

Live operation also needs an incident handling process: who reviews the alerts, within what time, and when a match counts as an incident. An alert is not yet a confirmed incident. Under Article 33 of the GDPR, a controller that becomes aware of a personal data breach must notify the supervisory authority without undue delay and, where feasible, within 72 hours, unless the breach is unlikely to result in a risk to the rights and freedoms of the data subjects. That is why the data protection officer also needs to know about the process. Behind the alerts there is often haste or a mistake, less often intent.

Quick self-test: pick three documents from your own folder and ask two colleagues which level they would assign them to. If the answers differ, the scheme or the written handling rules are not yet clear enough to build a DLP rule on.

Related service: Data loss prevention (DLP) →

Contact

Ready to secure your critical systems?

Our experts will assess where your security stands and build a roadmap that fits your organization.

ARLITECH's office in Gödöllő

Visit us

You are welcome to visit our Gödöllő office in person.

Address
Centrum Office Building, 2100 Gödöllő, Dózsa György út 13, 3rd floor, office 309
Office hours
Mon–Fri, 9 a.m. to 5 p.m.
Get directions (opens in a new tab)