Reducing Computational Waste and Untrustworthiness in Neural Network Training
Faculty Mentor Information
Dr. Tim Andersen (Mentor), Boise State University; and Dr. Gaby Dagher (Mentor), Boise State University
Presentation Date
7-2024
Abstract
In the field of artificial intelligence, neural networks have quickly become the standard in developing model architectures. As this burgeoning technological ecosystem continues to evolve, many industries stand to benefit from adopting such systems. Consequently, the integrity and reliability of these networks is paramount, especially in applications where security is critical. To this end, minimizing computational waste due to phenomena such as data poisoning is crucial. We introduce a novel blockchain-based protocol designed to enhance the security of neural network training. This protocol aims to detect and mitigate the effects of poisoning attacks, ensuring a trustworthy and reliable end model. Achieved through leveraging blockchain consensus technology and Merkle tree techniques, a transparent, immutable, and decentralized training process is introduced. Experiments demonstrating the protocol's efficacy in identifying and neutralizing poisoning attempts during training are conducted, demonstrably improving the overall integrity and reliability of the model's end state.
Reducing Computational Waste and Untrustworthiness in Neural Network Training
In the field of artificial intelligence, neural networks have quickly become the standard in developing model architectures. As this burgeoning technological ecosystem continues to evolve, many industries stand to benefit from adopting such systems. Consequently, the integrity and reliability of these networks is paramount, especially in applications where security is critical. To this end, minimizing computational waste due to phenomena such as data poisoning is crucial. We introduce a novel blockchain-based protocol designed to enhance the security of neural network training. This protocol aims to detect and mitigate the effects of poisoning attacks, ensuring a trustworthy and reliable end model. Achieved through leveraging blockchain consensus technology and Merkle tree techniques, a transparent, immutable, and decentralized training process is introduced. Experiments demonstrating the protocol's efficacy in identifying and neutralizing poisoning attempts during training are conducted, demonstrably improving the overall integrity and reliability of the model's end state.