Abstract: Blockchains based on Proof-of-Work can maintain a distributed ledger with a high security guarantee but also lead to severe energy waste due to the useless hash calculation. Proof-of-Useful-Work (PoUW) mechanisms are alternatives, but finding hard puzzles with easy verification and useful results is challenging. Recent popular deep learning algorithms require large amount of computation resources due to the large-scale training datasets and the complexity of the models. The work of deep learning training is useful, and the model verification process is much shorter than its training process. Therefore, in this paper, we propose DLchain, a PoUW-based blockchain using deep learning training as the hard puzzle. Theoretical analysis shows that \ours can achieve a security level comparable to existing PoW-based cryptocurrency when the miners' best interest is to maximize their revenue. Notably, this is achieved without relying on common assumptions made in existing PoUW-based blockchain such as globally synchronized timestamps.
Authors: Changhao Chenli, Boyang Li and Taeho Jung (University of Notre Dame, USA)
Email: cchenli@nd.edu, bli1@nd.edu, tjung@nd.edu