PoT consensus mechanism

In addition, we designed a PoT (Proof Of Training) consensus mechanism for model training and operation.

The PoT consensus mechanism is divided into the following stages:

• Initialize: Each node publishes the model and data they want to train to the network.

• Assign tasks: Tasks are assigned to various nodes in the network according to a certain strategy (for example, random or resource-based). The historical performance and reputation of nodes are taken into account when assigning tasks. For example, if a node has shown high efficiency and honesty in past training tasks, it may be prioritized in future task allocations. Efficient nodes should be given priority to AI models that run in real time. There is no guarantee that nodes required by the operating environment will only be used for model training.

• Train: Each node uses their computing resources to start a training task, while recording key training metrics (e.g., loss per iteration, GPU utilization, etc.) in a secure log. In order to prevent nodes from tampering with training indicators, zero-knowledge proof technology is used to verify the correctness of the training process.