2

I would like to ask, what are appropriate ways to adjust thresholds for a biometric system when we have only training and testing data of a single user? I am using Speaker, face and gait recognition but in mobile device scenario. In mobile scenario we know that we have always 1:1 matching case (We can go for 1:n matching but this would make system less usable as it is time taking, and need bigger data sets and we don't need it.). So, in this situation I want to find some way where I can find threshold from the training data. and whenever i get some test data I assume that this data is coming from genuine user and after matching, threshold basically decided if a that person is genuine/impostor.

  • Can you be more specific about what type of biometrics you're using? The term "biometrics" is broad and could refer to visual (like iris, fingerprint, vein pattern, etc), behavioural (like gait, speech, etc), odour (breath analysis), blood analysis. Also: what type of training algorithms are you using? What type of thresholding are you using? – Mike Ounsworth Jun 23 '15 at 13:42
  • I am using Speaker, face and gait recognition but in mobile device scenario. In mobile scenario we know that we have always 1:1 matching case. So, in this situation I want to find some way where I can find threshold from the training data. and whenever i get some test data I assume that this data is coming from genuine user and after matching threshold basically decided if a that person is genuine/impostor. – Muhammad Jun 23 '15 at 13:50
1

It depends on how the learning is done by the biometrics system.

If the system is distributed to you with prior knowledge of what a human face is, and if you're only teaching it the faces it should accept, it would be enough to just enrol the authorised users.

If the system is merely a computer vision learning model without any prior knowledge, you'd need to build a dataset of human faces representative of the human population, label the faces of the authorised users as authorised and the others as denied, and train the system.

I've never used the system you're asking about, so if you cannot figure out in which of those two categories it falls, I suggest you contact the customer support of the manufacturer.

1

Without training the system with data from other individuals, the concept of threshold is largely pointless.

When learning, all biometric systems need to be adjusted for false reject rate and false accept rate (FRR and FAR), but in your situation training will not give you any FAR data.

So although you can very easily make sure your system will always recognise your valid user, you have no idea whether it will accept no others, some who look like them or all humans (or possibly dogs, ostriches...who knows)

You need to be able to allow the system to determine the differences between your valid user and non-valid users, and then work out whether you can cope with some FRR (your user is carrying something heavy, perhaps) or with some FAR (someone with very similar gait and matching beard, say)

In summary - you can't do this unless you get a data set from many individuals

  • What about one-class learning? Don't have the time to create an answer right now, but I'd think that would be a possible approach (not that it is the best one). – Michael Jun 25 '15 at 15:26
  • Single class doesn't work for the requirements stated by the op – Rory Alsop Jun 25 '15 at 15:28
  • So, lets think for the fingerprint system that we have in laptops now a days, is it trained with data from other individuals too (I am not sure) and are shipped to the user with other users trained data? When we have 1:1 scenario, test data is only and only matched with claimed user model. and then a threshold is used to tell if test data is from genuine user or and Impostor. Face recognition application on mobile phones. Here I think they have only 1:1 matching, (Means matching with trained-data from genuine user only). Then how do they recognise without training with other user data – Muhammad Jun 26 '15 at 11:16
  • Same as @SteveDL's post - you have the existing data set shipped with it. – Rory Alsop Jun 26 '15 at 11:24
  • Even if we use FRR and FAR to determine a global threshold for a specific biometric, and if any template from same biometric when matched either under 1:1 or 1:n scenario has similarity above that threshold or distance below that global threshold will be accepted either its cat/dog face. – Muhammad Jun 26 '15 at 11:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.