Features

Solai meets all the features mentioned in the introduction :

Infringement on Data Privacy

Solai trains the global model not by collecting personal data like the existing AI training scheme but by letting the participants train their own local model and just aggregating them. For more security, we hide the local models in a veil using a cryptographic technique called Differential Privacy (which will be explained later in this whitepaper)

Get Reward and Ownership

Local models are issued in the form of NFT on a chain, and its trainer can own it. And the contributors get rewards whenever the global model is used. This is possible because of the transparency and tamper-proof of blockchain smart contracts(programs in Solana). And also sustainable by the following :

• The trainer commits its result on a blockchain meaning anyone can verify it. This would lead to the correct rewarding system among Federated Learning participants because they would participate in the system only when they think it is valid.

• Since the commits on the blockchain are tamper-proof, they can be tokenized and utilized as a means of incentive. We can also construct a token economy that distributes the incoming fee whenever the global model is used.

Robustness to Malicious

Participants Malicious actions can be detected and stopped in Solai. We will describe the method in the next part.