Biometric modality: Face – key considerations
Regarding occlusions by glasses, masks etc:
While face recognition accuracy was historically impacted when glasses, hairs or other elements occluded features of the face, current face recognition algorithms are resilient to most occlusions as long as they do not cumulate. Sunglasses should still be preferably avoided as they cover the most valuable part of the face.
The COVID epidemic has caused the algorithms to evolve and adapt. Face recognition, when the capture subject wears a sanitary mask, is now at an accuracy level similar to that achieved by the best face recognition algorithms a few years ago.
Regarding so-called bias, demographic differentials:
Face recognition algorithms are based on machine learning that relies on a training dataset. Poorly designed algorithms and training datasets with unbalanced distribution of demographic factors (like age, gender or ethnicity) may result in a difference of accuracy between members of various demographic groups. Face recognition algorithms should be properly evaluated to ensure they have a homogenous accuracy between various groups.
One-to-one (1:1) verification applications
The process of authenticating an individual’s identity by comparing face images can be conducted manually i.e. a human adjudicator assesses the two images and decides if they are of the same person or, alternatively, the process can be automated so that if the face recognition system generates a satisfactory similarity score between the two presented face images it can allow actions (e.g. open an access gate) or transactions (e.g. permit digital onboarding through a device or online).
One-to-many (1:N) identification applications
Face images may be regarded as less intrusive than, for example, fingerprints as face images can be obtained without direct or close contact with the individual. However, this also means that these images may be obtained without the individual’s knowledge or consent. The fact that face images can be captured at a distance or from online sources raises other privacy concerns in modern society. The accuracy of one-to-many (1:N) identification face recognition systems has also been questioned in terms of demographic differentials i.e. the ability of the software to match persons from different genders or skin tones based on the training data used in developing the system. Recent testing has shown that the best algorithms do not exhibit demographic differentials to any significant degree but the performance of some other algorithms are less favourable in this respect. (Refer to the Biometrics Institute Good Practice Framework Row B – Societal Impact and the Biometrics Institute Three Laws of Biometrics – formulate policy then processes and then the technology).
The use of face recognition systems in public places by law enforcement agencies has been challenged in several countries from a civil rights perspective and this has led to the authorities banning the use of the technology in some jurisdictions and some suppliers of the systems invoking voluntary moratoria while legislation catches up with this fast-evolving technology.
The accuracy of one-to-many (1:N) identification face recognition applications has improved to the point where most use cases can be fully automated as for one-to-one (1:1) face verification systems. However, for some specific use cases, such as the evaluation of low quality images in law enforcement applications, human adjudication is usually required to review the suggested candidate(s) generated by the face recognition search through the database/watch list (Refer to the Biometrics Institute Good Practice Framework A4.4/5/6). One-to-many (1:N) identification face recognition applications typically return a list of these potential match candidates to a human adjudicator, based on a predetermined number of high similarity scoring images, a similarity score threshold, or a combination of the two. The adjudicator then makes a final decision. As a consequence, the training and competence of human face recognition examiners is being reviewed and formalised in many countries and jurisdictions in order to reassure the public and the courts that face recognition is being applied on a sound scientific and impartial basis.
