What is Edge AI?

Edge AI - definition

Edge AI refers to running machine learning model inference directly on edge devices or in their immediate proximity - such as cameras, video recorders, workstations, on‑premise servers, or IoT gateways - instead of in the cloud. The primary goals are to reduce latency, limit data transfer, and minimize the exposure of personal data outside a controlled environment. In the literature and technical standards, Edge AI is often described as part of the broader edge computing ecosystem (e.g. ETSI MEC, ISO/IEC TR 23188), where local computing resources deliver services close to the data sources.

In the context of image and video anonymization, Edge AI means that the detection of sensitive objects (such as faces and license plates) and the application of blurring or masking techniques are performed locally. This approach shortens the processing path and reduces the risk of data leakage during transmission to the cloud. Deep learning models are trained in a development environment and then deployed for inference at the edge of the network.

The role of Edge AI in image and video anonymization

Visual anonymization first requires reliable detection of the areas that need to be concealed. In practice, the processing pipeline includes detecting and optionally tracking faces and license plates, followed by applying blur or pixelation filters. Running these steps at the edge eliminates the need to upload full‑frame footage to the cloud and makes it easier to comply with the GDPR data minimization principle. Edge AI is essential here, because without effective deep learning-based detection models, it is not possible to automatically and consistently identify regions that must be anonymized.

In Gallio PRO, detection and masking are performed locally in an on‑premise environment. The software automates the blurring of faces and license plates, but it does not automatically detect other objects (such as logos, tattoos, or documents) - these can be masked manually in the editor. Gallio PRO does not perform real‑time anonymization of video streams; processing is batch‑based.

Edge AI technologies and architecture for face and license plate blurring

Deployments rely on proven hardware-software stacks and models optimized for inference. Below are example components and best practices.

  • Models: Object detectors trained on datasets dedicated to faces (e.g. WIDER FACE) and license plates; single‑stage architectures (YOLO, SSD) or specialized models (RetinaFace). Full body segmentation is not required, as only faces and plates are masked.
  • Optimization: Conversion to ONNX, compilation with TensorRT or OpenVINO, post‑training INT8 quantization, and structured pruning - to reduce latency and model size without significant quality loss (see NVIDIA TensorRT and Intel OpenVINO documentation).
  • Runtime: ONNX Runtime, TensorRT, OpenVINO, TFLite, Core ML - depending on the platform.
  • Hardware: Data‑center or desktop‑class GPUs, VPU/NPU accelerators, and embedded platforms (e.g. NVIDIA Jetson, Intel iGPU with AI acceleration such as VNNI instructions), providing MAC acceleration and hardware video decoding.
  • Input/Output: Hardware video decoding, tiled processing for high resolutions, and storage of detection metadata and masks without retaining raw biometric features.

Key Edge AI parameters and metrics in anonymization

Assessing the quality and security of edge processing should be based on model metrics and operational parameters. The table below summarizes the most important measures.

Attribute

Description

Relevance for the DPO

 

Precision / Recall

Accuracy and completeness of face and license plate detection on test datasets

High recall reduces the risk of data exposure due to missed objects

mAP

Mean Average Precision across different IoU thresholds

Model comparability and quality regression tracking between versions

Masking IoU threshold

Coverage threshold of the detection area by the mask

Control over the safety margin of the mask

Inference latency

Processing time per frame or batch of frames

Planning offline processing windows

Throughput

Frames per second for given hardware and resolution

Estimating batch processing time and SLAs

Memory footprint

Model size and RAM/VRAM usage during inference

On‑premise platform selection and task segmentation

Power consumption

Energy usage under inference load

Operational costs and ESG considerations

Domain robustness

Stability across lighting conditions, angles, and resolutions

Risk of false negatives in atypical scenes

Benefits of using Edge AI for data protection

Edge processing supports data minimization and integrity while ensuring operational reliability in demanding environments. Commonly highlighted benefits include:

  • Reduced transfer of personal data to the cloud - smaller attack surface and easier access control.
  • Deterministic latency - predictable batch queues and schedules.
  • Connectivity independence - offline processing in isolated network segments.
  • Stronger compliance with the privacy by design principle - limiting the scale and retention time of raw visual data.

Challenges and limitations of Edge AI in anonymization

Edge AI requires careful model lifecycle management and risk control. Key challenges include:

  • False negatives in difficult conditions - an undetected face or license plate is a potential confidentiality breach.
  • Domain drift - changes in cameras, lighting, or scenes can degrade performance without model updates.
  • Resource constraints - memory, thermal limits, and power budgets affect architecture and input resolution choices.
  • Validation and auditability - the need for versioning models, test datasets, and mask configurations to demonstrate due diligence.

Use cases and best practices

Edge AI is used in environments where footage must not leave controlled infrastructure and where data transfer costs are significant. In practice:

  • Anonymizing CCTV recordings before sharing footage for access requests or training purposes.
  • Blurring faces and license plates in images from road tests, inspections, and technical documentation.
  • Batch processing archives on on‑premise servers, allocating compute resources outside peak hours.
  • Configuring mask margins and quality control with output sampling and system logs that do not store personal data - in line with data minimization principles.

Standards and references

Definitions and practices should align with recognized standards and data protection guidelines.

  • ETSI GS MEC 001 V2.2.1, Multi‑access Edge Computing - Terminology, 2018.
  • ISO/IEC TR 23188:2020, Cloud computing - Edge computing landscape.
  • Regulation (EU) 2016/679 (GDPR), Recital 26 and Article 4 - definitions of personal data and anonymization criteria.
  • EDPB, Guidelines 3/2019 on processing of personal data through video devices, final version 2020.
  • ISO/IEC 20889:2018, Privacy enhancing data de‑identification - terminology and classification.
  • NVIDIA, TensorRT Developer Guide - INT8 quantization and inference optimization.
  • Intel, OpenVINO Toolkit Documentation - model optimization and edge deployment.

Implementation notes in Gallio PRO

In Gallio PRO, Edge AI is used for local detection and blurring of faces and license plates in photo and video materials. The system operates on‑premise and does not perform real‑time stream anonymization. Other image elements, such as logos, tattoos, or documents, can be blurred manually by the user in the editor. The software does not collect logs containing face or license plate detection data and does not store personal or sensitive data. This approach supports minimal data transfer and compliance with privacy by design principles.