How to encrypt the extracted biometric features at the output of classifier/neural network training?

Question

Neural networks

Good day! A question arose regarding biometric identification, namely the case when a user can be authorized by a photo (simulated). But the snag is the following - I need to get at the output of the classifier / neural network not a class label, but already a key based on the features extracted from the face image (the usual "recognition" by a face from a webcam is a stage that I have long passed and implemented).
I myself am now digging into arrays of foreign English-language articles (and this is a very long task), so I decided to ask a question here, maybe someone has already worked in this area and will suggest methods, or some options, so that you can immediately read on topic, not so much :)
PS I found an option about fuzzy extractors while I'm studying the "nature" of this method :)

Answer 1

Train an autoencoder on your data. Then take the middle layer (the smallest in terms of the number of neurons), it will be the "abstract" code of your image. Or I didn't understand the question?

Answer 2

Regarding security, I clearly described everything in the comments to my article.
If you check biometric features using an MLP classifier, then it makes no sense to encrypt, the neural network weight matrix itself will already be something like a hash, and the feed-forward of the input vector through the neural network will be reconciled with this hash. Thus, according to my method, we have N weight matrices, where N is the number of users, it no longer makes sense to store vectors.
https://habrahabr.ru/post/336198/
By the way, luckily, just 5 days after your question. Read the article, you can pick up interesting ideas from there.