Towards Smart Contracts for Verifying DNN Model Generation Process with the Blockchain

In recent years, DNNs (Deep Neural Networks) have been applied into various fields and expected to be deployed into real-world applications. On the other hand, lack of transparency in DNNs makes them unreliable. To ensure transparency in DNN models, it's necessary that model validators can verify the entire learning process and convince third parties with limited resource that the given model is correctly generated. For this purpose, we propose a smart contract that is based on the dispute resolution protocol for verifying DNN model generation process. We divide the entire learning process into layer-based computations. The necessary data for validating each computation, such as the outputs of neurons, weights between layers and their gradients, are uniquely determined by the one-way hash function and the hashes are combined by multiple structured Merkle trees. This enables an honest validator to make a proof that asserts the target model is incorrectly generated, and third parties can check whether the assertion is true by only performing the given computation. Finally, to reveal how our proposal affects the performance degradation during the training, we evaluated running time for deep learning that enables our proofs. This result shows that our proposal can be applied into real applications. For this purpose, we propose a smart contract that is based on the dispute resolution protocol for verifying DNN model generation process. We divide the entire learning process into layer-based computations. The necessary data for validating each computation, such as the outputs of neurons, weights between layers and their gradients, are uniquely determined by the one-way hash function and the hashes are combined by multiple structured Merkle trees. This enables an honest validator to make a proof that asserts the target model is incorrectly generated, and third parties can check whether the assertion is true by only performing the given computation. Finally, to reveal how our proposal affects the performance degradation during the training, we evaluated running time for deep learning that enables our proofs. This result shows that our proposal can be applied into real applications. Finally, to reveal how our proposal affects the performance degradation during the training, we evaluated running time for deep learning that enables our proofs. This result shows that our proposal can be applied into real applications.