AI zero-knowledge learning

Zero-knowledge machine learning (ZKML) is an emerging technology that combines zero-knowledge proofs (ZKP) and machine learning (ML). ZKP is a cryptographic protocol that allows one party (the prover) to prove to another party (the verifier) that they know some information without revealing any details about that information. In the context of AI machine learning, ZKML can be used to protect the privacy of training data and models.

Real-time monitoring: All nodes send key training indicators during their training process to other nodes in the network. This can be achieved through secure communication channels. In addition to key training indicators, other behaviors of nodes can also be monitored, such as network latency, computing resource usage, etc., to more comprehensively evaluate the performance of nodes. Score the node's historical performance and reputation.

Probabilistic Epoch Validation (PEV): At the end of each epoch, each node verifies its performance by randomly selecting a portion of the data and recalculating the training loss for these samples, and then comparing it with the training loss for the entire epoch. correctness of training. When validating training losses, statistical tests can be used to detect the presence of anomalies. For example, if a node's training loss is much lower than other nodes, then that node may need further review.

Verifiable Model Ownership (VMO): Each node needs to store the model state or checkpoint during its training process. After training is complete, these checkpoints can be used to prove that the node actually trained the model. To prevent nodes from forging model checkpoints, nodes can be required to generate a unique identifier (such as a hash value) associated with the checkpoint when saving the checkpoint. This identifier can then be used to verify the authenticity of the model checkpoint.

Validation and consensus: After training is completed, all nodes share their training results and logs. All nodes then verify that the training process of other nodes is correct based on the shared information. If a node's training process is verified to be incorrect or dishonest, the node's results will be ignored. Only when a majority of nodes reach a consensus that a node's training process is correct, the node's results will be accepted.

Update model: Once consensus is reached, all nodes will update their models with the accepted results. Then, the process returns to step 2 to start the next training task. When updating the model, you can consider using some advanced model fusion techniques, such as federated learning, to improve the performance and accuracy of the model.