Verifiable SD-JWT Credential on Mobile

We previously measured the performance - in terms of proof verification - of our Noir circuits, including a verifiable SD-JWT credential as used in SWIYU on commodity hardware, namely a MacBook Pro.

While at first glance it appears to be only the next logical step to also measure the performance of our circuits - including, most importantly, the verifiable SD-JWT credential - it is in reality a major milestone in our efforts to add zero-knowledge proofs to the Swiss e-ID ecosystem.

The existing privacy-preserving digital credentials, Crescent and Longfellow which we discussed in a previous blog post, rely on specialised credentials to achieve their impressive speed of < 1s proof verification time. We, on the other hand, are able to use the standard SD-JWT through a Noir circuit, a major prerequisite for integrating zero-knowledge proofs into the Swiss e-ID.

Using the Mopro toolkit, we created a simple Android app to execute and benchmark our different circuits. Despite using a somewhat dated Galaxy A54 5G, we are very happy with the results.

Noir test	Average of 100 runs [s]	Best [s]	Worst [s]
c03_holder_binding	1.975	1.841	2.176
c04_issuer_signature	2.428	2.073	2.821
c05_age_verification	0.314	0.258	0.381
c06_non_revocation	2.645	2.109	3.016
c09_full_proof	6.483	5.643	6.940
d10_swiyu_jwt	18.419	16.154	21.013

Note: refer to zkp-pocs for circuit implementation details.

While at first sight the runtime of the verifiable SD-JWT credential does seem to leave much to be desired, it is important to point out that this is an entirely unoptimised circuit. Its runtime of 18.5 seconds on a low-end mobile phone is unarguably too slow for an agreeable user experience. However, we are confident that the new prover backend we are currently working on will bring the runtime well below 10 seconds.

Beyond the performance measurements themselves it is a major step to have successfully generated a SD-JWT credential proof. This moved our endeavour from a theoretical undertaking to a practically applicable result.

The next steps are to confirm our performance measurements using a second library, Noir Android, and to continue working on optimising the Noir circuits to achieve better performance.