Skip to content

[4.0.0] - Release Notes

This release updates SelphID SDK with new features related to Facial Authentication, Facial Identification and Facial Liveness evaluation processes. This release is recommended for customers who are employing Facial Recognition and Liveness Evaluation and require major feasibility and usage in their projects.

New facial biometric engine implemented.

A new biometric engine for Facial Recognition employed for face comparison, related to Facial Authentication or Facial Identification processes is integrated in this release. This new biometric engine improves:

  • Security
  • Usability
  • Onboarding use cases (selfie vs document).

Template Compatibility

Due to this new biometric engine, the Facial Templates used in processes of identification and gallery creation are NOT compatible with previous versions. Owing to that, a new Facial Template generation for 1:N processes must be performed from images or from raw facial templates.

Evaluation set

A test set has been used that contains experiments focused on Facial Recognition contemplating the following scenarios:

  • Good quality
  • Poor quality
  • Lighting changes
  • User pose
  • On Boarding
  • Bias(Racial)
  • Masks

Population: 55M comparisons. The experiments are carried out trying to reproduce the scenarios and circumstances that can happen in production projects.

Metrics

To carry out the study on the operation of the new SDK, the following metrics have been taken into account according to ISO 19795-1, which are used in the NIST evaluations:

  • FTE (failure-to-enrol): It is the proportion of failed attempts in the generation of biometric patterns. Failures can happen because the software throws an exception or refuses to process the input image. This usually happens if there is no face or if the image quality is extremely poor.
  • FNMR(false non-match rate): Ratio employed to decide that two biometric patterns are not of the same identity, while in reality they are. This metric does not count FTEs as an error.
  • FMR (false match rate): Proportion employed to decide that two biometric patterns are of the same identity, while in reality they are not. This metric does not count FTEs as an error. In addition to these metrics defined in ISO 19795-1, the employment of the FNMR@FMR metric is made, which allows to measure how much the system incorrectly rejects a set of specific security threshold. In NIST evaluations this value is variable depending on the type of the database evaluated, however, for this report it has been chosen the typical value of FNMR@FMR = 0.1%, which corresponds to the percentage of false rejection when the system accepts wrong 1 out of every 1000 enrollment attempts.

New facial passive liveness engine implemented

A new facial passive liveness employed for facial spoof detection related to evaluate the liveness from a single image is integrated in this release. This new facial passive liveness engine improves:

  • Security
  • Usability

Evaluation set

A test set has been used that contains experiments focused on Facial Spoof Detection contemplating the following scenarios:

  • Genuines
  • Paper
  • Phone
  • Tablet
  • Screen
  • Screen 4K
  • Photo
  • Paper mask
  • Latex mask

Population: 382K samples The experiments carried out are based on different circumstances described in the evaluations of ISO 30107.

Metrics

To carry out the study on the operation of the new SDK, the following metrics have been taken into account according to ISO 30107:

  • FTA: It is the proportion of samples that could not be acquired because the expected biometric characteristic was not detected compared to the total number of attempts made.
  • APCER: It is the proportion of Presentation Attacks (Attacks) incorrectly classified as genuine against the total number of attempts made in a specific scenario.
  • BPCER: It is the ratio of Bona Fide Presentations (Genuine) incorrectly classified as Presentation Attacks against the total number of attempts made in a specific scenario.

Fixed bugs and changes

  • Restored Uncertain value from FacialAuthenticationStatus and FacialLivenessDiagnostic.
  • Restored FrontalFast value from FacialDetectionType.