Processing Latency - definition
Processing latency is the total time from receiving input data to producing a finished processing output. In the context of image and video anonymization, it refers to the time from reading a file or frame, through face and license plate detection, applying blur masks, to saving and optionally recompressing the output. In standards and academic literature, processing latency is part of system performance quality characteristics, defined in ISO/IEC 25010 as time behavior (response times, processing times, throughput) [ISO/IEC 25010:2011]. For online services, network delay as defined in ITU‑T Y.1541 may be a component, but in batch image/video anonymization, computational and I/O latency usually dominates.
In practice, both end-to-end latency and stage-level latency are measured to identify bottlenecks and ensure compliance with legal and operational requirements (e.g., deadlines for delivering anonymized materials).
The role of processing latency in image and video anonymization
Face and license plate anonymization requires object detection using deep learning models. Such models are essential for automatically locating areas to be blurred during processing. Processing latency directly affects how quickly a secure, anonymized version of a file becomes available. This is important for accountability and for designing workflows aligned with the privacy by design principle under Article 25 of the GDPR, as well as with the EDPB Guidelines 3/2019 on video devices (Version 2.0, 29 January 2020).
Gallio PRO operates in batch mode, does not perform real-time anonymization, and automatically blurs only faces and license plates. Therefore, processing latency is defined here as the batch duration from the start of processing to the availability of the output file. Logs do not contain detection details or personal data.
Technologies and architecture affecting latency
Latency is influenced by hardware, software, and encoding layers. Below are the main components of the processing pipeline with typical sources of delay.
Component | Description | Typical sources of latency
|
|---|---|---|
Input / I/O | File reading, serialization, disk-to-memory transfer | Disk throughput, network, container format |
Decoding | Decoding H.264/H.265, JPEG | GOP structures, B-frames, lack of hardware acceleration [ITU‑T H.264, H.265] |
Preprocessing | Scaling, normalization, color space conversion | Memory operations, lack of vector optimization |
Inference | Face / license plate detection using neural networks | Model size and complexity, lack of accelerators, batch size |
Postprocessing | Applying masks, frame merging, NMS | Algorithmic complexity, CPU/GPU operations |
Output encoding | Image/video compression and saving | Bitrate parameters, profile, encoder preset |
An on‑premise architecture reduces dependency on network latency and the risk of transferring data outside the organization. GPU acceleration (e.g., CUDA), inference compilers (e.g., TensorRT, OpenVINO), and efficient I/O libraries reduce inference and transformation times without changing the semantics of the anonymization process.
Key parameters and metrics
Latency should be reported using metrics that are robust to variance and expressed in units appropriate to the task scale. The following attributes are useful for DPOs and technical teams.
Metric | Definition | Unit
|
|---|---|---|
Total latency L | Time from batch start to completed output file | s, min, h |
Per-frame latency | Average time to process a single frame | ms/frame |
p95 / p99 percentiles | Latency not exceeded by 95% / 99% of jobs | ms, s |
Throughput | Number of frames or minutes of video processed per unit of time | FPS, min/h |
Delivery time SLA | Guaranteed deadline for completing a batch | h, days |
A reference latency decomposition for a single frame is: Lframe = tio_in + tdecode + tpre + tinf + tpost + tencode + tio_out. For a batch of N frames with parallelism P: Lbatch ≈ ceil(N/P) × median(Lframe) + queue overhead.
Measurement and reporting
Measurement should be deterministic and repeatable, with time synchronization when the pipeline spans multiple nodes. For precise timestamping in distributed systems, IEEE 1588‑2019 Precision Time Protocol is used. In integrated on‑premise environments, monotonic system clocks and consistent telemetry are sufficient.
- Stage instrumentation: timestamps before and after each pipeline step, aggregated into p95/p99 reports.
- Stable inputs: fixed encoder preset, identical models and batch size, disabled turbo/boost modes during benchmark tests.
- DPO reports: batch duration, throughput, percentage of jobs meeting SLA, variance between batches.
Challenges and limitations
Reducing processing latency must not compromise anonymization quality. A lighter model may speed up inference but reduce face or license plate detection sensitivity. Output compression increases encoding latency but reduces file size. Gallio PRO does not perform real-time anonymization, so optimizations focus on batch processing time and deterministic results. The system does not automatically detect logos, tattoos, name badges, or monitor screens; these elements can be blurred manually in the editor, introducing an additional, controlled operational latency.
Use cases
In practice, DPOs and security teams use processing latency to plan processing windows and assess risk. Typical examples include anonymizing CCTV footage before disclosure, preparing training datasets with blurred faces, and batch processing event photos where legal exemptions do not apply. In many Western European countries, blurring license plates is common due to identification risks and local practices, adding extra inference load to the pipeline. In Poland, interpretations of license plate status vary, but guidance from the Polish DPA (UODO), the EDPB, and CJEU case law tends to support anonymization, which should be considered when planning SLAs.
Standards and references
The following standards and guidelines define time metrics, synchronization, and the legal framework for processing video containing personal data.
- ISO/IEC 25010:2011 - Systems and software quality models. The performance efficiency section defines time behavior, including processing times and throughput. https://www.iso.org/standard/35733.html
- ITU‑T Y.1541 (2011) - Network performance objectives for IP‑based services. Defines network latency classes. https://www.itu.int/rec/T-REC-Y.1541
- IEEE 1588‑2019 - Precision Time Protocol. Time synchronization standard for distributed measurements. https://standards.ieee.org/standard/1588-2019.html
- ITU‑T H.264 | ISO/IEC 14496‑10 AVC and ITU‑T H.265 | ISO/IEC 23008‑2 HEVC - Video encoding standards affecting decoding and encoding latency. https://www.itu.int/rec/T-REC-H.264 and https://www.itu.int/rec/T-REC-H.265
- GDPR - Regulation (EU) 2016/679, Article 25 privacy by design and by default. https://eur-lex.europa.eu/eli/reg/2016/679/oj
- EDPB, Guidelines 3/2019 on processing of personal data through video devices, Version 2.0, 29 January 2020. https://edpb.europa.eu