Decentralization + ZK + blockchain + AI training.
This approach combines cutting-edge fields to facilitate AI training. It proposes a method for collaborative AI training within a decentralized network, utilizing Zero-Knowledge (ZK) proofs to ensure privacy.
The challenge with conventional methods is that participants must disclose their data; furthermore, they might possess low-quality data or act dishonestly. The proposed approach addresses these issues through several mechanisms:
Verification: A participant runs a consensus-agreed AI model on their own data and generates a ZK proof confirming the correct execution of the computation, which the network then verifies.
Filtering: Validators analyze the gradients used to update model parameters and create "masks." In this context, a mask acts as a quality control mechanism, specifying which parameters to retain fully and which to attenuate.
arxiv.org/abs/2611.11477
Terminus needs JavaScript to show replies and pictures.