MindLayer: Proof of Intelligence

Unlike Proof of Working (PoW) and Proof of Stacking (PoS), which rely on solving complex mathematical problems or holding a certain amount of tokens, Proof of Intelligence (PoI) leverages the power of machine learning (ML) tasks. In other words, nodes within the AI network are rewarded for performing ML tasks instead of traditional methods. However, to safeguard the results of machine learning tasks and mitigate potential threats from malicious FHE validators, Mind Network proposes MindLayer. This entails calculating rewards and reaching a consensus based on FHE-encrypted data. By leveraging FHE, sensitive data remains encrypted throughout the process, ensuring confidentiality and integrity.

In MindLayer, four clusters orchestrate the validators: the miners cluster, validators cluster, consensus cluster, and rewards cluster. The miners' cluster plays a critical role in the model generation, while validators assess these models for accuracy. These clusters, namely miners and validators, constitute essential components of the subnet, collectively driving the network's functionality. Conversely, the consensus cluster is entrusted with the task of achieving consensus among diverse validators, ensuring agreement on critical decisions within the network. Similarly, the rewards cluster operates after the consensus cluster, focusing on calculating rewards allocated to the Mind subnet participants. Together, these clusters form the backbone of the Mind Validation Service, strengthening the network's infrastructure and fostering its operational efficiency. The intricate process of MindLayer, fortified by FHE, unfolds as follows:

• Initialize public keys and parameters: The rewards cluster initiates the process by generating a public key $pk$and a secret key $sk$ using FHE $Gen()$ function. Subsequently, the public key $pk$ is disseminated for encryption purposes, while the secret key $sk$ is divided into $l$ pieces, named as $(sk_1,sk_2 ...sk_n)$, and distributed among reward nodes for the threshold decryption later.

• Model generation: Miners undertake machine learning tasks, generating results pivotal for subsequent evaluations, denoted as $(m_1,m2,...,m_j)$, where $j$ is the number of miners.

• Validation: Validators meticulously evaluate the results produced by different miners, encapsulating their assessments within encrypted ciphertexts utilizing the FHE public key. For each validator with its staking amount $S_i$, a corresponding weight vector $w_j$ is generated, representing its validation on miner $j$. To protect their validation results, validators use the public key $pk$ to encrypt their weight values. Since validators do not have access to the secret keys, it is impossible for them to decrypt the encrypted data and gain knowledge of other validators' validations. All the validators will send their encrypted weights to the consensus cluster.

• Consensus evaluation: The Consensus cluster receives the encrypted vector from all validators, resulting in a weight matrix $w_{i,j}$that represents the validation results of validator $i$ for miner $j$. Together with the staking vector, the encrypted weighted matrix $W$are seamlessly transmitted to the consensus cluster for FHE evaluation. This phase encompasses the calculation of trust, consensus rates, and rewards under the veil of ciphertext. To enhance the integrity of the consensus process, a designated consensus node spearheads the dissemination and conclusion of results, supported by specialized consensus nodes proficient in precise FHE calculations. Detailed insights into these calculations can be referred to the whitepaper.

• Rewards decryption: Upon the completion of encrypted rankings and rewards calculation, the rewards cluster receives the results. Leveraging threshold FHE, it initiates the decryption process, meticulously decrypting a portion of the encrypted values. Each reward node is equipped with a portion of secret keys $sk_1,...,sk_l$, where $l$ represents the number of reward nodes. With their respective partial secret keys, each node can decrypt a part of the plaintext. The decryption stops until the established threshold is attained, and the subsequent transmission of plaintext results to the smart contract, thus culminating in the completion of the transaction cycle.