Develop an FHE-based iris identification application tutorial with Concrete

[gelintonx]

Overview

Design an iris recognition biometric template protection schemes based on Homomorphic Encryption

Description

Biometric recognition is becoming a prominent way to authenticate users or verify their identities. As highlighted in the ISO/IEC 24745, it is important to protect biometric information for secrecy, irreversibility, and renewability during storage and transfer. In this bounty you will need to design an FHE based remote authentication system that protects sensitive Iris information during storage and biometric comparison. In its paper "Hybrid biometric template protection: Resolving the agony of choice between bloom filters and homomorphic encryption", the research team looked at three different approaches: Bloom filters, homomorphic encryption and hybrid biometric template protection (BTP). The team highlighted the advantages and disadvantages of each approach.

The bounty objective is to:

• use Concrete to implement a single key TFHE-based BTP for an access control system
• all reference templates are stored encrypted in a database on the server

The client:

• collects the iris biometric (format should be the same as UBIRIS.v2 or CASIA)
• extracts the feature vector
• encrypt it
• and send it to the server

Then the server:

• perform the comparison against its database
• send an encrypted matching ID back to client
• a no match encrypted message is returned if no matching template is found

Performance:

• Good comparison performance
• Possibility to implement multi threading

Your PR should comply with the following:

• Create the app frontends/concrete-python/examples/iris-identification/{client,server}.py
• Create the tutorial docs/tutorial/iris-identification.{md,ipynb}

Library targeted

• Concrete

Reward

Up to €5,000

Related links and references

• Concrete documentation
• Resolving the agony of choice between bloom filters and homomorphic encryption
• UBIRIS.v2 Use password: UBIRIS2_IEEETPAMI_101109_200966
• CASIA

[bcm-at-zama]

Hello @gelintonx

Some comments on https://github.com/zama-ai/bounty-GacsBiometrics/tree/main (as it appeared yesterday, ie

commit 1630873d6e1e70fec2f5dc972b04958e4b2a565d Date: Mon Dec 18 00:04:19 2023 +0100

beta v1.4

):

• python circuit/fhe_circuit.py does not work:

  it says FileNotFoundError: [Errno 2] No such file or directory: '../template-database/C1_S1_I1.mat' I have tried to find . -name C1_S1_I1.mat but nothing; certainly the file name has changed

• in your README:

  • could we have more story telling, to explain what you're doing here, what's Iris check, why there is a need for privacy etc; it would be great that when someone lands on your repository, we have a clear view of what's going on, including some references to Concrete ML
  • could we have an example of a command line call with a match, and a second call with a non-match? It would be great it's clear for users how it works when we have the right Iris and when we have not the right one
  • also, I would say few words about the distance you use, and how you implemented it in FHE (possibly with few words about the difference successive improvements, and their speed)

• maybe it's possible to make your program a bit faster, by replacing

    inp = 2 * probe_template + registered_template
    xor = fhe.univariate(lambda x: (x & 1) ^ ((x >> 1) & 1))(inp)

by

    inp = probe_template + registered_template
    xor = fhe.univariate(lambda x: x % 2)(inp)

It will compute XOR as well, I have tested on my computer (a 2019 Intel MacBookPro) and it seems a bit faster. Maybe to check on your side as well? I would say the bitwidth of the input of the PBS is as large (2 bits) but maybe there is less noise so the PBS can be faster. I did 5 inferences:

    old method: [82.257, 88.515, 100.945, 97.644, 93.514] seconds
    new method: [77.174, 79.680, 87.305, 91.242, 93.006] seconds

• I did not try the client / server for now

• Also, you could / should link to your docs/Project flow.md from the main README, like "Read the engineering documentation here"

• also, stricty speaking, do you follow the

Create the app frontends/concrete-python/examples/iris-identification/{client,server}.py
Create the tutorial docs/tutorial/iris-identification.{md,ipynb}

of the open bounty submission? In particular, having a notebook is really a plus for people understanding.

In a nutshell, I would work on the packaging and documentation, to be more explicit to users.

Congrats for what you've done. Cheers

[gelintonx]

Dear @bcm-at-zama thank you for this review of the project. I have taken note of everything to repair it as soon as possible. If there is anything else I can help with I am available. Again, thank you very much.

[bcm-at-zama]

Thanks a lot, @gelintonx . You'll tell me if something is unclear. Maybe just keep in mind that it's one of the reviews, you should have one more after the break. Cheers

[gelintonx]

@bcm-at-zama Good morning and Happy new year! I have been testing a new template created using University of Salzburg kit it's size is 256 x 1 but the result is different from using 20 x 480 template. Performance different is huge 1.676 vs 68.335 seconds I would like to know your opinion about this difference. Also I added some new features and will commit soon.

[bcm-at-zama]

Happy new year. How should I reproduce the problem, please?

[gelintonx]

Good morning @bcm-at-zama this is the project I said https://wavelab.at/sources/USIT/. Templates are smaller than GacsBiometrics templates.

[bcm-at-zama]

Could you provide sources files and instructions on how to reproduce the problem, please? certainly the best is to create a GitHub issue in https://github.com/zama-ai/concrete/issues and someone in the team will look at it. If it's more convenient, don't hesitate to create a branch or repo somewhere with everything which is needed to reproduce the problem.

Thanks a lot

[bcm-at-zama]

And in the meantime, do you think we could already merge what you've done (with the comments I had given)? I really would love your biometry work to be public, such that we can show it to Concrete users

[gelintonx]

@bcm-at-zama Yes, I would also love to make it public. I have to push a new commit that I have been scheduling since the beginning of January.

[bcm-at-zama]

Hey @gelintonx . Let's try to finish this, can we?

[gelintonx]

@bcm-at-zama Sure I have been really busy but I also want to finish the project.

[bcm-at-zama]

We take the lead here from now, we are going to reward you for the work, thanks a lot for your contribution (https://github.com/zama-ai/bounty-GacsBiometrics).

[gelintonx]

@bcm-at-zama I get it, it's been a long time since I started. Thanks Zama team it has been a great opportunity to work with FHE