Appendix A: Problem Solving

If you have an issue with facial recognition, use the following flow chart to identify the type of issue you are facing, and what to do about it:

‍

Are Faces Being Detected Reliably?

To answer the question ‘Are faces being detected reliably?’, you need to touch the home screen so that the Welcome screen displays. There is a small red progress bar at the top of the video. If this bar quickly progresses from left to right, and the Enter button is enabled, then a face has been detected. If this red bar does not progress, and the Enter button remains disabled, then a face has not been detected.

If you are having issues with an individual, you should ask that individual to stand in front of the iPad and touch the home screen so that the Welcome screen is displayed. If you stand to the side, you will be able to see whether their face is being detected or not by watching the red progress bar and Enter button. Repeat that process a few times, and you should have a good understanding as to what is happening.

‍

Liveness Testing – Real People Being Detected as Fraud

To verify whether you have a Liveness Testing issue, turn the Visual Feedback switch on.

The welcome screen will then flash red when NoahFace believes it has detected fraud. If you determine you definitely have a Liveness Testing issue (ie: you are regularly seeing the screen flash red when real people present themselves), you should consider the following:

• Firstly, if you have multiple liveness testing algorithms enabled, you should try each algorithm one by one to see which one is causing your issues.
• Is it possible to move anything in the video background (eg: shelving, a picture on the wall, etc)?
• Is it possible to move the iPad to another location with a different background?
• For Access Control solutions, is it possible to angle the iPad slightly (ie: so the left or right vertical edge is further from the wall than the opposite edge). This often changes the camera view sufficiently so that a door frame or window is no longer in view.
• For solutions such as Time and Attendance, is it possible to mount the iPad lower and angle it towards the ceiling (at up to 45 degrees)? This often changes the camera view sufficiently so that a picture, shelving, or a window is no longer in view.
• Is it possible to zoom the image slightly so that a door frame or window is no longer in the video? Note that zooming the video will slow down the recognition process significantly, so this should be done only as a last resort.

‍

Generally, one of the above approaches will work, and will allow you to use liveness testing in almost all environments. However, if you have exhausted all options with both the Movement Analysis and Constant Background algorithms, you should consider the Depth Analysis algorithm. This algorithm is not sensitive to the background at all, and will work well in almost all environments.

‍

Liveness Testing – Fraud not Detected

If someone has conducted fraud (eg: presented a selfie of someone else which has been recognised as a real person), and you want to try to prevent this happening in the future, you should do the following:

1. Open up the person’s user details (in the NoahFace App), and press the right-most toolbar button to view their biometrics. Remove their consent (press the “-” icon) which will cause their biometrics to be deleted from all iPads that are currently active.

‍

It may take up to 5 minutes for the deletions to replicate to other iPads. If you are using Cloud based biometric storage, you should wait for this to complete before proceeding.

2. Ensure the iPad is properly mounted. The liveness testing algorithms will NOT work reliably unless the iPad is properly mounted in a fixed location.
3. Re-Read the section on Liveness Detection, and review and adjust your Liveness Testing settings as appropriate.
4. Re-Read the section on Biometric Learning, and consider setting this to Registration .In this mode learning will only take place during the registration process, and certain registration methods (eg: List+Admin) require that a supervisor is present (when fraud is generally easy to prevent). Under most circumstances we recommend leaving this setting at its default of Continuous (as this results in much more reliable recognition), this is a trade-off you can consider if fraud detection is critical to your application.
5. When you re-start the App, step to the side and wait for the message “Preparing liveness testing....” to disappear. If you start using the App before this message disappears, the liveness testing algorithms will NOT work.
6. You are now ready to re-register staff.

‍

Note that the basic Liveness Testing algorithms (eg: Movement Analysis) will detect many, but not all, attempts at fraud. If fraud detection is absolutely critical in your environment, we strongly recommend considering purchasing an iPad Pro and using the Depth Analysis liveness test. This is by far the strongest fraud detection algorithm.

You should also consider other, non-technology based, measures – eg: surveillance cameras, employment contract terms, etc. These play a critical role in preventing fraud.

‍

Detection Issues (False Negatives)

Detection issues can occur on individuals that are extremely tall or short. In such cases, your only real options are:

• Increase the Detection Distance setting.
• Make these individuals aware they are out of frame, so they crouch or stand on their toes as needed.
• Ask these individuals to stand further from the iPad.

‍

Detection issues can also occur on people with extremely dark skin (particularly if lighting is poor) or with extremely thick glasses. In such cases, your options are:

• Increase the Brightness/ISO setting (and re-testing with people with fair skin).
• Decrease the Detection Threshold setting (to Low or Very Low).
• Ask an individual to take of their glasses to be recognised.
• Improve the image quality by adding shading or lighting.

‍

Incorrect Matches (False Positives)

If an individual experiences an incorrect match, it is intended that they should press the ‘Not Me’ button. This will automatically unlearn the association between their user record and the biometric (or biometrics) that caused the incorrect match.

‍

If users do not press the ‘Not Me’ button, and instead complete a transaction (such as clocking in) as the wrong user, their biometrics will be recorded against the matched user.

‍

See: ‘Appendix B: What do I do if I have an incorrect match?’ for more information.

‍