@AlleManfredi@SuccinctLabs@zama Reminds me what we did a year ago with Private DAO at Zama Hackathon https://t.co/dxqpJTk8dZ .
Key is that you need to remember that for voting HE is usually enough https://t.co/qt8C9hzMZN
@Ingo_zk Really cool work, did not know about the problem before hand! One question - if you run a model on CPU only, no multithreading, would it still be non-deterministic?
I understand that you loose in inference time, however it is anyway magnitudes quicker than actual proving
@AlexanderJLong What about security? I found that this is a biggest caviate for decentralized training:
1. Replication factor (or cryptography overhead) you need to verify that the results are correct
2. Ability to inject backdoors for adversarial parties (similar to Federated Learning attacks)
@ilblackdragon Considering attacks such as data reordering (https://t.co/yumG0H2XgX), even making data public will do little to confirm lack of backdoors.
Furthermore, backdoors are just a subset of adversarial attacks, so rather more effort is needed in making models "robust" overall
@_Daniel_Ospina What if the personalized AI agents can not just choose an option, but participate in decisions making on your behalf? Saving each of us a lot of time and making decisions more involved
Exciting times for the ZKML space as Zero Gravity continues its work exploring and developing the use of zero-knowledge proofs for Weightless Neural Networks (WNNs) .
Read the blog post by @elus_aegis and @georgwiese to learn more!
https://t.co/HJ1tQPMmuq