Skip to content

Pipelines

PAD pipelines

PAD pipelines overview

The set of algorithms the IDLive Face relies on is grouped together in an entity called a pipeline. We constantly improve the algorithms, and each new set is bundled as a separate pipeline.

  • pad-r-1. Added in version 1.47. This is the default pipeline. This standard option offers a 25% improvement in reducing false rejections (BPCER) and a 30% improvement in reducing false acceptances (APCER). It provides robust performance with optimized processing speed and accuracy.
  • pad-c-1. Added in version 1.47. Available as a separate Docker container, this pipeline brings advanced calibration options. Soft calibration: 2.5x improvement in reducing false rejections (BPCER), Hardened calibration: 3x improvement in reducing false acceptances (APCER), Regular calibration: 1.8x improvement in both BPCER and APCER. This enhanced pipeline offers better accuracy but requires more processing time, making it ideal as a secondary check when the regular pipeline flags an image for review.
  • Perseus. Added in version 1.44. This pipeline demonstrates better accuracy on live samples captured in low light conditions. It also shows a significant improvement on high quality 3D masks. The soft calibration of this pipeline shows a very low BPCER without significantly affecting APCER.
  • Iris. Added in version 1.42. The Iris pipeline improves accuracy regarding fabric masks and screen replays. It operates comparably with Astraea in terms of performance and improves preciseness in Liveness detection.
  • Hestia. Added in version 1.41. The Hestia pipeline significantly enhances APCER and BPCER metrics, particularly in the mobile domain. It operates approximately twice as slowly as Astraea, making it ideal for scenarios where accuracy is prioritized over speed. Explicit enablement is required to activate this pipeline, and future updates are expected to optimize its processing speed.
  • Pegasus. Added in version 1.39. Pegasus, replacing the Heron pipeline, offers improved performance with lower latency and ten times the accuracy. This pipeline is well-suited for applications demanding high performance and speed, such as time-sensitive or high-volume tasks.
  • Astraea. Added in version 1.37, removed in version 1.44.
  • Aphrodite. Added in version 1.36, removed in version 1.42.
  • Apollo. Added in version 1.35, removed in version 1.40.
  • Heron. Added in version 1.16, removed in version 1.39.
  • Artemis. Added in version 1.31, removed in version 1.37.
  • Theia. Added in version 1.29, removed in version 1.37.
  • Dionysus. Added in version 1.27, removed in version 1.35.
  • Hephaestus. Added in version 1.24, removed in version 1.31.
  • Zeus. Added in version 1.21, removed in version 1.29.
  • Jupiter. Added in version 1.18, removed in version 1.23.

DFD pipelines

DFD pipelines overview

The set of algorithms the IDLive Face Deepfake Detection relies on is grouped together in an entity called a pipeline. We constantly improve the algorithms, and each new set is bundled as a separate pipeline.

  • dfd-1. Added in version 1.45. This pipeline provides robust protection against various deepfakes and digitally manipulated images. It can identify AI faceswaps, including complex deepfakes, and synthetic images created by tools like Midjourney and Stable Diffusion. The system also detects digital modifications such as watermarks, borders, screenshots, or overlays. While it currently provides basic protection against face reenactment and audio-driven deepfakes, such as lipsyncing or talking heads, improvements in these areas are expected in future updates for enhanced security.

Choosing pipeline

By default IDLive Face Server uses the latest available pipeline. To select another one put its name into the IDFACE_SERVER_PIPELINE environment variable:

docker run --env IDFACE_SERVER_PIPELINE=apollo ...

You can load additional pipelines by setting the IDFACE_SERVER_AVAILABLE_PIPELINES environment variable. Several pipelines should be separated with a comma:

docker run \
    --env IDFACE_SERVER_PIPELINE=apollo \
    --env IDFACE_SERVER_AVAILABLE_PIPELINES='astraea,apollo'
    ...

Note that each additional pipeline will increase the IDLive Face Server's memory usage and a startup time.

Pipeline aliases

In addition to specifying pipeline names explicitly, pipeline aliases are also available and do not change from release to release. While the actual underlying pipelines might change, these aliases can simplify continuous migration by always invoking the most recent liveness pipeline for a given attack vector.

Currently available pipeline aliases:

  • default-pad alias for pad-r-1
  • default-dfd alias for dfd-1

Aliases can be specified instead of pipeline names on server configuration:

docker run \
    --env IDFACE_SERVER_PIPELINE=default-pad \
    --env IDFACE_SERVER_AVAILABLE_PIPELINES='default-pad,default-dfd,hestia'
    ...

And also pipeline aliases can be specified instead of pipeline names in REST API calls e.g. for /check_image:

'data' parameter structure
{
  "domain": "GENERAL",
  "pipelines": [
    {
      "pipeline": "default-pad",
      "calibration": "HARDENED"
    },
    {
      "pipeline": "default-dfd",
      "calibration": "REGULAR"
    },
    {
      "pipeline": "hestia",
      "calibration": "REGULAR"
    }
  ]
}