Artificial Intelligence in Data Security - Definition
Artificial intelligence in data security refers to the use of machine learning and deep learning techniques to identify, classify, process, and protect information in order to reduce the risk of breaches affecting confidentiality, integrity, and availability. In the context of image and video anonymization, this means automatically detecting elements that enable personal identification - particularly faces and license plates - and then distorting them (e.g., through blurring or pixelation) in line with the principles of data minimization and privacy by design (Articles 5 and 25 of the GDPR - Regulation (EU) 2016/679).
The definition and framework for the secure use of AI are further supported by ISO/IEC standards on AI risk management (ISO/IEC 23894:2023) and information security management systems (ISO/IEC 27001:2022).
The Role of AI in Image and Video Anonymization
In face and license plate blurring, AI acts as both the detection and decision-making layer. Deep learning models trained on datasets of faces and vehicle registration plates recognize these objects under varying conditions, after which transformation algorithms apply blur or pixelation effects. This requires models trained on representative datasets - otherwise detection may be incomplete, directly increasing the risk of personal data disclosure.
In practice:
- object detectors locate faces and license plates using convolutional neural networks (CNNs) or vision transformer models,
- object tracking mechanisms stabilize masks across video sequences,
- image processing filters introduce pixel-level distortion.
In on-premise solutions such as Gallio PRO, AI automates face and license plate anonymization. Other elements (e.g., logos, tattoos, name badges, documents) can be blurred manually in the editor. This division can minimize the scope of data processing and simplify risk assessment in accordance with the privacy by design principle (Article 25 GDPR) and the EDPB Guidelines 3/2019 on video devices (final version 2020).
Technologies and Architectures (Computer Vision)
AI-based video anonymization software relies on established computer vision paradigms. The choice of technology affects detection sensitivity and speed, and indirectly the residual risk of data exposure. Commonly used approaches include:
- Object detection: RetinaFace (Deng et al., 2019) for face detection; the YOLO family and its variants for license plate detection - returning bounding boxes and confidence scores (AP/mAP as standard metrics; WIDER FACE benchmark for face detection),
- Segmentation/landmarks: precise face masks improve coverage in cases of partial occlusion,
- Video tracking: algorithms such as SORT/DeepSORT stabilize masks between frames, reducing “flickering,”
- Image processing: Gaussian blur, pixelation, or replacing a region with a neutral color - the chosen method determines resistance to reconstruction and contextual readability,
- On-premise MLOps: model version control, change validation, reproducible builds, and environment isolation (aligned with ISO/IEC 27001:2022 requirements for change control and secure development environments).
Key Parameters and Metrics (Visual Anonymization)
The effectiveness and operational security of AI-powered data anonymization must be measured and documented. The table below summarizes common metrics used in AI projects for image and video anonymization. Threshold values depend on the specific risk context and should result from a formal risk assessment (ISO/IEC 23894:2023; NIST AI RMF 1.0, 2023).
Metric | Description | Implementation Notes | Source/Standard
|
|---|---|---|---|
Precision/Recall (Detection) | Ratio of correct detections and detection completeness | Calculated per class (face, license plate); IoU and confidence thresholds must be specified | WIDER FACE, COCO metrics |
mAP@IoU | Mean Average Precision at a defined Intersection over Union threshold | Report IoU value (e.g., 0.5) and PR curves | WIDER FACE benchmark; COCO evaluation |
False Negative Rate | Percentage of undetected objects | Critical for personal data exposure risk | NIST AI RMF 1.0 (2023) |
Latency | Processing time per frame/image | Specify hardware configuration and batch size | ISO/IEC 25010 (performance efficiency) |
Mask Stability | Percentage of frames without mask “flicker” | Requires sequential testing with tracking enabled | Engineering practice; NIST SP 800-53 rev.5 (testing and evaluation controls) |
Resistance to Reconstruction | Verification that applied transformations prevent re-identification | Reconstruction attack testing and risk estimation required | ISO/IEC 20889:2018 (privacy-enhancing techniques terminology) |
Challenges and Limitations (Risk and Compliance)
AI in image anonymization involves both technical and legal risks. Variable recording conditions, compression artifacts, occlusions, and long-distance shots increase the likelihood of missing parts of faces or license plates. Models may lose generalization outside their training domain, requiring continuous validation and retraining.
From a compliance perspective, the EDPB emphasizes that a person’s image may constitute personal data, and video processing must comply with GDPR principles, including necessity and proportionality assessments (EDPB Guidelines 3/2019, final version 29 January 2020). In Poland, legal interpretations and case law regarding license plates as personal data may vary, while in many Western European countries supervisory authorities more consistently expect license plates to be masked before publication or sharing. These differences should be reflected in risk assessments and publication policies.
Practical Use Cases (Face and License Plate Anonymization)
AI-based video anonymization is widely used in repetitive processes where performance and consistency are critical. Below are typical on-premise deployment scenarios (excluding real-time or live stream anonymization):
- CCTV archives: bulk face and license plate blurring before sharing footage with third parties (EDPB 3/2019),
- Research and audits: preparing video and image datasets for publication in reports in accordance with the data minimization principle,
- Media and communications: publishing event footage while respecting image rights exemptions and masking license plates in line with EU regulatory practices,
- Data subject access requests: selectively blurring bystanders in materials provided to applicants.
In this context, Gallio PRO automatically blurs faces and license plates, while other elements are handled manually in the editor. The software operates fully on-premise and does not perform real-time or live-stream anonymization. The system does not store logs containing detected faces, license plates, or personal data.
Standards and Regulatory References
Proper use of artificial intelligence in data security for image and video processing should rely on recognized standards and authoritative guidance. The following references provide essential benchmarks for system design, risk assessment, and documentation:
- GDPR - Regulation (EU) 2016/679: Article 5 (principles), Article 25 (privacy by design), Recital 26 (anonymization - “information which does not relate to an identified or identifiable natural person”),
- EDPB Guidelines 3/2019 on processing personal data through video devices (final version 29 January 2020),
- ISO/IEC 27001:2022 - Information Security Management System (ISMS), including controls for secure development and maintenance,
- ISO/IEC 23894:2023 - AI risk management, including identification of technical and ethical risks,
- ISO/IEC 20889:2018 - Privacy-enhancing data de-identification techniques,
- ISO/IEC 22989:2022 and ISO/IEC 23053:2022 - AI terminology and system lifecycle,
- NIST AI Risk Management Framework 1.0 (2023) - model risk identification, assessment, and monitoring,
- NIST SP 800-53 rev. 5 (2020) - security and privacy controls for information systems, including testing and validation requirements.