Aim: We are currently building a decentralized ML training prototype on a p2p networking stack, compatible with mobile devices and modern deep learning libraries.
Join us: You're very welcome to join the conversation on our slack space, as well as on the issues pages here.
Motivation: Machine learning applications have recently gained remarkable impact on our everyday lives. However, significant concerns and risks arise as such systems are operated in a centralized way: Currently, most machine learning models important on a societal level do not have their utility, control and privacy aligned with the data ownership of the participants. This highlights the need for new mechanisms for collaborative training while respecting data privacy. While federated learning is a first step towards this goal, decentralized training goes further and puts users in control of the entire collaborative training process, without any central point of failure.
In this project we aim to build decentralized analogues of current machine learning algorithms, which are at the same time (i) efficient (R1), (ii) privacy-preserving (R2), (iii) robust to malicious actors (R3), and (iv) support individual consent and transparency on the resulting utility of trained ML models. We follow a public models, private data approach.