What Is On-Premise Processing?

On-Premise Processing - Definition

On-premise processing (also commonly referred to as on-premises processing) means performing data operations within infrastructure owned and controlled by an organization, without permanently transferring the data to a public cloud. In the context of image and video anonymization, this involves running detection, classification, and masking of faces and license plates on local servers - often within networks isolated from the Internet.

An on-premise approach supports the principles of data minimization, integrity, and confidentiality under Article 5(1)(c) and (f) of the GDPR, as well as the implementation of appropriate technical and organizational measures under Article 32. It reduces the data transfer surface, simplifies access control, and helps meet data residency requirements.

The Role of On-Premise Processing in Image and Video Anonymization

Image and video anonymization requires detecting sensitive objects and irreversibly distorting them. In practice, deep neural networks trained on labeled datasets are used and then deployed for inference during the masking process. While model training typically takes place outside the customer’s environment, inference and the entire processing pipeline can be conducted on-premise to ensure that source data never leaves the controlled infrastructure.

In solutions such as Gallio PRO, automatic blurring applies to faces and license plates. The system does not blur entire silhouettes, does not provide real-time anonymization, and does not support live video stream anonymization. Other visual elements - such as logos, tattoos, ID badges, or screen content - can be manually blurred using the built-in editor. The software does not collect logs containing images or other personal data; in particular, operational logs should not store metadata that could identify individuals.

On-Premise Technologies and Architecture

An on-premise implementation for video anonymization software focuses on a cohesive architecture covering compute, storage, and security layers. Below are the most common components and deployment best practices.

• Compute layer - GPU-enabled servers for deep learning inference; alternatively CPUs with SIMD acceleration. Containerization (e.g., Docker) and orchestration in isolated environments.
• Data layer - Local file or object storage with encryption at rest (e.g., AES-256; key length and cryptoperiod selection in line with NIST SP 800-57), along with retention control and versioning.
• Transport security - TLS 1.2 or 1.3 (RFC 5246 and RFC 8446) for internal traffic, with network access restrictions via ACLs and segmentation.
• Identity and access management - Integration with AD/LDAP, least privilege principle, and separation of operator and administrator roles.
• ML lifecycle management - Receipt of signed inference models, integrity validation, controlled offline updates, and regression testing of detection quality before deployment.
• Operational audit trail - Logging of administrative activities without storing image content or personally identifiable metadata.

Key Parameters and Metrics (On-Premise Anonymization)

Evaluating an on-premise processing solution should consider both detection quality and operational characteristics. The table below outlines key metrics and how they are verified.

Challenges and Limitations

While on-premise processing increases control, it also places full responsibility for maintenance and updates on the organization. The main risks that must be addressed through architecture and procedures include:

• Capital and energy costs of infrastructure and GPUs.
• Managing security patches and container images in isolated networks.
• Batch scalability for large video archives and diverse codecs.
• Model drift and the need for periodic detection quality validation.
• Log and telemetry control to prevent storage of personal data.
• Supply chain security for models and libraries, including signature verification and SBOM validation.

Use Cases

On-premise video anonymization is preferred where full processing control and zero cloud transfer are required. Typical use cases include public sector entities, law enforcement agencies, and industrial organizations.

• Sharing CCTV footage with third parties after blurring faces and license plates.
• Publishing workplace training materials with individuals anonymized.
• Internal health and safety (HSE) incident analysis while preserving employee privacy.
• Handling evidentiary materials requiring chain of custody and integrity assurance.

Normative References and Sources

The following documents provide the definitional and technical foundation for on-premise processing in the context of data protection and security measures used in visual anonymization.

Legal note: In many jurisdictions, facial images and license plates may constitute personal data under the GDPR. Specific obligations and exemptions depend on national law and the context of publication.