This means that if the son initially fails to unlock his parent’s device, but then the password is entered whilst in view of the sensor, his Face ID data can be brought into the neural network’s processing. This would make it more likely for his face to unlock the device in future, even if the initial setup process was the mother alone.
In a WIRED article interviewing the mother, they claim that this was not what had happened. However, it’s honestly the most likely explanation for how this would work and it’s very easy to accidentally train it without thinking. The lack of defined facial features for the child will also play a role here.
The article also suggests that the lightning conditions of the initial training had a noticeable effect on the accuracy in this particular case:
At WIRED’s suggestion, Malik asked his wife to re-register her face to see what would happen. After Sherwani freshly programmed her face into the phone, it no longer allowed Ammar access. To further test it, Sherwani tried registering her face again a few hours later, to replicate the indoor, nighttime lighting conditions in which she first set up her iPhone X. The problem returned; Ammar unlocked the phone on his third try this time. It worked again on his sixth try. At that point, Malik says, the phone’s AI seemed to learn Ammar’s features, and he could consistently unlock it again and again.
Touch ID included much of the same learning mechanisms as Face ID but the difference between the two is that genetics do not make it more likely for people you know (your family) to have similar fingerprints as you. In contrast, it is much more likely for a member of your family to look similar to you and confuse the Face ID learning process.
The training process for Face ID only kicks in if the face data matches to a ‘certain threshold’. What Apple may do in a future software update is increase this threshold of likeness. This would reduce the number of false-positives for the training to consider, making it harder for face data from family members to contribute to the learning process.
The downside of doing this is that Face ID would take longer to learn about situations where it really is you unlocking it, but fails to recognise you.
Another possible way Apple could improve the reliability of Face ID for people with similar-looking family members is to offer an ‘advanced training mode’ in Face ID settings. Face ID setup process only asks for two scans of a person’s face.
An additional training mode would allow users to volunteer more ‘trusted’ information to the system, improving the neural network models with more data. This would reduce the chance of incorrect matches.
More Info: 9to5mac.com