Abstract Title
Verifying Integrity of Neural Networks via Blockchain
Abstract
Our research introduces a unique blockchain network that verifies the integrity of extensively trained neural network models. It works towards ensuring data corruption or manipulation during model training did not occur, offering a tamper-proof ledger for tracking the verified evolution and documentation of models. The system involves two main parties: Submitters, providing trained models with complete training records, and Verifiers, who re-compute slices of the training records. The blockchain system strikes a balance between computational power and responsibility by selectively retraining significant intervals rather than the entire model, thus saving computation resources. An algorithm employing a unique method of model-analyzation is utilized to spot anomalies within distinct iterations of the training of the model, thus identifying components to recompute. This meticulous verification system embodies trust, democratization, and computational efficiency, marking a significant progress in neural network applications across various domains, while maintaining confidence and trust in their outcomes.
Verifying Integrity of Neural Networks via Blockchain
Our research introduces a unique blockchain network that verifies the integrity of extensively trained neural network models. It works towards ensuring data corruption or manipulation during model training did not occur, offering a tamper-proof ledger for tracking the verified evolution and documentation of models. The system involves two main parties: Submitters, providing trained models with complete training records, and Verifiers, who re-compute slices of the training records. The blockchain system strikes a balance between computational power and responsibility by selectively retraining significant intervals rather than the entire model, thus saving computation resources. An algorithm employing a unique method of model-analyzation is utilized to spot anomalies within distinct iterations of the training of the model, thus identifying components to recompute. This meticulous verification system embodies trust, democratization, and computational efficiency, marking a significant progress in neural network applications across various domains, while maintaining confidence and trust in their outcomes.