University of Notre Dame
Browse

Proof-of-Deep-Learning Consensus in Blockchain Systems

Download (8.75 MB)
thesis
posted on 2022-12-07, 00:00 authored by Boyang Li

Bitcoin, as one of the most successful blockchain applications, attracts unprecedented attention and investment from both academia and industry. To maintain the consistency of transaction data, the proof-of-work consensus mechanism utilizes the brute-force algorithm for validation and implicitly hosts a competition of hardware and energy. To mine cryptocurrency more efficiently, people developed GPU mining, FPGA mining, and ASIC machine mining. However, these methods still suffer from the issue of wasting energy because the hashing algorithm has remained unchanged. Alternatively, the proof-of-useful-work consensus has been proposed to execute relatively more useful tasks to maintain consistency, therefore the ``wasted energy'' could contribute to useful work.

In the proof-of-deep-learning consensus, for the first time, I proposed the two-phase design which utilized the deep learning training process as the workload of blockchain. To further improve the consensus, I enhanced the feasibility to exploit the computation power of blockchain for deep learning algorithms. The miners can switch between different training tasks in the memory pool and align the target task.

However at this stage, all miners are still required to work on the same task in the same block, as a result, the majority of energy has been wasted due to the computation redundancy. Therefore, I introduced the collaboration strategy to reduce redundancy.

In the proof-of-federated-learning-subchain work, I categorized different miner roles and emphasized the importance of data contributors. Under this framework, miners can select partners based on average response time and training data value to form pools under a federated learning framework.

In the series of my work, I aim to recycle the "waste" resources to train deep learning models which powered many applications behind the scenes in our daily lives. Beginning with the initial proof-of-deep-learning consensus, I continued to improve feasibility, reduce redundancy, and initialize the framework of the next-generation consensus which is optimized for the deep learning training tasks as the workload of blockchain.

History

Date Modified

2022-12-21

Defense Date

2022-07-07

CIP Code

  • 40.0501

Research Director(s)

Yiyu Shi

Committee Members

Taeho Jung Walter Scheirer Jinjun Xiong

Degree

  • Doctor of Philosophy

Degree Level

  • Doctoral Dissertation

Alternate Identifier

1355547104

OCLC Number

1355547104

Additional Groups

  • Computer Science and Engineering

Program Name

  • Computer Science and Engineering

Usage metrics

    Dissertations

    Categories

    No categories selected

    Keywords

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC