KVKK Limits in Video Analytics: Face Recognition and Biometric Data

Why is face recognition different under KVKK? The difference between face detection and recognition, biometric templates, data risk by type of analytics and privacy controls at installation.
Technical summary
Technology
Video analytics, face detection, face recognition, re-ID, metadata, KVKK Law No. 6698
Topic
The personal and biometric data nature of analytics outputs
Level
Intermediate
Prerequisites
  • A list of the analytics to be used
  • The existing KVKK inventory for the camera system
  • Support from a data protection advisor or the legal unit
In this article7 sections

In video analytics projects, the KVKK conversation usually gets stuck on the question “are camera recordings personal data?” The real issue lies elsewhere: analytics produces new types of data alongside video. Counts, events, attribute labels, track information and, in some products, face templates. Below we’ve gathered, for technical teams, which outputs open the discussion of personal, biometric or special category data, the technical difference between face detection and face recognition, and design options that lower risk at the installation stage. The content isn’t legal advice; make implementation decisions together with your data protection advisor and legal unit.

Analytics outputs and personal data

Law No. 6698 on the Protection of Personal Data defines personal data as any information relating to an identified or identifiable natural person. The text of the Law is available on mevzuat.gov.tr. For the technical team, the key word in the definition is “identifiable”: even if an analytics output doesn’t contain the person’s name, it can be personal data if it can be combined with other information and linked to the person.

No name doesn’t mean no person.

It helps to think of the outputs analytics systems produce in four groups:

  • Raw and processed video: Recordings, event clips, thumbnails. If a person is visible, it’s personal data.
  • Event records: Records such as “Camera 7, 14:32, zone intrusion.” On their own they don’t identify a person, but together with the linked event clip they lead to one.
  • Per-person metadata: Track number, bounding boxes, clothing color, bag, direction of movement, plate text. Combined with time and place information, it can point to a person. We explained how this data is used in search in the smart video search with metadata article.
  • Aggregated statistics: Hourly entry counts, average occupancy. If no per-person record is kept, they can’t be linked to a person.

The goal in design is simple: settle for output from the lowest group needed to achieve the purpose. An occupancy report doesn’t need per-person metadata, and an intrusion alarm doesn’t need a face template. Counting a person, detecting that a face is present in the image and recognizing the person are separate functions. The result being an aggregate count doesn’t mean the source image is anonymous; state separately whether raw video, metadata and any biometric template are kept.

The difference between face detection and face recognition

The two terms are constantly confused in sales documents. Their technical and legal consequences, however, are different.

  • Face detection: Finds the region of the image containing a face and produces box coordinates. It doesn’t determine who it is. It’s used for tasks such as face blurring, counting faces or choosing the best face frame.
  • Face template extraction: From the detected face, a numerical vector (template, embedding) representing the person’s unique facial features is generated. Biometric processing starts here.
  • Verification (1:1): The template is compared with a single registered template belonging to the identity the person claims; for example, whether the person swiping a card is the cardholder.
  • Identification (1:N): The template is compared with many templates in a database; a watchlist or a “who is this person?” search.

Don’t assume that a product using face detection doesn’t generate and store templates in the background. Some platforms store the best face frame and its template by default so that searches can be done later. In product evaluation, verify this behavior explicitly from the configuration screen and documentation; a sentence in a sales presentation isn’t enough.

Biometric data and special category data

Article 6 of KVKK lists biometric and genetic data as special category personal data, together with data such as health, religion, and attire and clothing. The conditions for processing this data are narrower than for general personal data, and Article 6 was amended in 2024 by Law No. 7499, which redefined the processing conditions. Follow the current text and the Authority’s related guidance and decisions on the Turkish Personal Data Protection Authority website. In terms of technical measures, the Board’s decision on “adequate measures” for special category data is also among the key reference texts.

The Law doesn’t separately define biometric data. The GDPR definition in Europe (Art. 4(14)) is technically instructive: data relating to the physical, physiological or behavioral characteristics of a person, resulting from specific technical processing, that allows or confirms the unique identification of that person. Under this definition, a photograph alone isn’t biometric data; a face template extracted from the same photograph and used for matching, however, is.

The difference lies in the technical processing.

In video analytics, two gray areas need particular attention:

  • Appearance-based re-identification (re-ID): For tracking people across cameras, a vector is generated from clothing and body appearance. This vector usually relies on temporary appearance rather than a permanent identity characteristic, but because it serves to track a person uniquely, it should be assessed separately. Methods based on behavioral characteristics such as gait are close to biometric recognition.
  • Attribute classifiers: Estimates such as age, gender or a headscarf can produce indirect information about special categories (for example, belief via attire and clothing). Turn off these attributes, whose contribution to incident review is limited.

Risk table by type of analytics

The table classifies, from a technical perspective, the nature of the data produced by common types of analytics. The risk level is given relatively, based on how linkable the data is to a person and its biometric nature; the final assessment is made with the project’s purpose and context.

Analytics Data produced Nature of the data Relative risk
People counting (counts only) Total counts Aggregated Low
Virtual fence / zone intrusion Events and clips Personal data Low to medium
PPE detection Events, clips, zones Personal data, linkable to employees Intermediate
Attribute search Per-person metadata Personal data Intermediate
License plate recognition Plate text, time, place Personal data Intermediate
Face detection / blurring Face position Personal data, no template Low to medium
Appearance-based re-ID Appearance vector Should be assessed separately Medium–high
Face recognition (1:1 or 1:N) Face template, match Biometric, special category High

With PPE and similar occupational safety analytics, there’s a drift we often see on site: the purpose is a safety warning, but over time the outputs are carried over into employee performance evaluation. Assess this boundary together with the framework in the employee monitoring with cameras in the workplace article.

Privacy-focused design options

Most of the risk is reduced at the product selection and configuration stage of enterprise camera systems. Rolling it back after installation is much harder. Technical options:

An architecture producing only numerical metadata

For purposes such as people counting, occupancy and density, analytics runs on the camera or a local device and sends only counts outside. Video isn’t recorded; if it’s recorded for a separate security purpose, it isn’t linked to the analytics output. Per-person track information is discarded after processing.

Face and person anonymization

Blurring faces or the whole silhouette of a person in live viewing or on export reduces the personal data load in the image, as long as incident review doesn’t require otherwise. We compared masking methods in the privacy masking in cameras article.

Avoiding or narrowing face recognition

If the purpose can be met by another method (access cards, plates, operator verification), don’t use face recognition. If its use is unavoidable, prefer 1:1 verification over 1:N identification; where possible, keep templates in an environment or device belonging to the user rather than a central database; keep watchlists limited to the purpose and time-limited.

Retention and access

Retention periods for metadata and event clips are defined together with video, and automatic deletion is applied. Search and export permissions are tied to separate roles, and every query is written to the audit trail. Template databases are encrypted and protected by separate permissions. A search screen without an audit trail is a monitoring tool that can’t be audited.

Checklist

This list can be used as technical input to the project’s privacy impact assessment work:

  • The purpose of each analytics is written down; the relationship between the purpose and the data produced is justified.
  • The outputs produced by each analytics (video, events, metadata, templates) are listed and classified according to the groups in the table.
  • Whether the product generates face templates or appearance vectors has been verified from documentation and configuration.
  • Unused attributes (age, gender, etc.) and search features have been turned off.
  • For purposes where aggregated data is enough, per-person metadata isn’t stored.
  • Retention periods and automatic deletion have been defined for metadata, event clips and templates.
  • Search, export and template management permissions have been given to separate roles; queries are written to the audit trail.
  • If biometric processing is planned, the legal basis, alternative method and adequate measures have been documented together with the legal unit.
  • The privacy notice and warning signs have been updated to cover the use of analytics.
  • If cloud-based analytics is used, where the data is processed and its transfer-abroad status have been determined.

Selecting and configuring analytics features according to these principles can be handled under AI video analytics deployment, and updating privacy notice and privacy processes together with our KVKK camera monitoring and privacy notice processes work.

Frequently asked questions

Is using face detection for face blurring processing biometric data?

Face detection only finds where a face is in the image; it doesn’t produce a template to identify the person and doesn’t make comparisons. That’s why on its own it’s assessed differently from face recognition. Even so, the image processed is personal data; verify technically whether detection results are recorded and whether the product generates templates in the background.

Does people counting analytics process personal data?

For counting, camera images are processed, and these images are personal data. If the design produces and stores only total counts, doesn’t record video and doesn’t keep per-person track information, the output becomes statistics that can’t be linked to a person. The risk being low depends on verifying that this design is really applied on site.

Can face recognition access control be installed with employees’ explicit consent?

This is a question requiring legal assessment. Technically, for explicit consent to rest on free will, it’s important that employees who don’t consent are offered an alternative access method such as a card, and that templates are kept separate, encrypted and with restricted access. Make the decision together with your data protection advisor and legal unit, according to current legislation.

When should a privacy impact assessment be done in a video analytics project?

Before product selection. Which analytics will be used, which outputs will be stored and who will have access directly determine product and architecture choice. An assessment done after purchase often turns into a study that just documents the existing installation.

Support

Contact us for support or a discovery session on this topic.

Contact form
[email protected]

Eretek Video Surveillance

IP cameras, recording and central management infrastructure. We plan blind spots, retention periods and the camera network around the risks of each site.

Knowledge Center

Other articles