ZK ML

What are you trying to do?We want to facilitate and incentivize large-scale cooperative machine learning while preserving individual data andmodel privacy. How is it done today? What are the limitations of the current system?Current federated learning uses privacy-enhancing techniques, but it is centralized and people rightly don’t trust it. What is new in your approach and why do you think it will be successful?The full realisation of our project would see the creation of a decentralised environment where autonomous and human actors would contribute to improve and leverage on various instances of artificial intelligence. If successful, what difference will it make?We have seen a growing number of initiatives deploying various forms of collaborative learning in the scientific community, thus highlighting the appeal of early adopters to mine more than just their own data sources to train more powerful and unbiased models. How long will it take?The mid-term will take 6-12 months while the full vision may require multiple years to complete. What are the mid-term and final exams to check for completeness?Verifiable evaluation of ML models (succinct zk proofs) will enable innovative use cases such as an Iterated prediction market (to establish a model’s value) and a Market for model queries (verified use of known models as a service.) The final product is fully decentralised model training. Model IP will be protected via either FHE or MPC. The owner obtains strong guarantees that the decentralised training was performed according to their preset conditions.Individual contributors and institutions are rewarded in accordance with the significance of their contribution.