DEEP CHAIN LEARNING BY COLLUSIONS OVER NETWORK WITH IMPROVED BLOCKCHAIN SECURITY

Sarkar, Ayush

Permalink

https://hdl.handle.net/2142/125060 Copy

Description

Title

DEEP CHAIN LEARNING BY COLLUSIONS OVER NETWORK WITH IMPROVED BLOCKCHAIN SECURITY

Author(s)

Sarkar, Ayush

Issue Date

2020-05-01

Keyword(s)

blockchain; smart contracts; deep learning; security

Abstract

Smart contracts in blockchains can be executed for identifying and verifying images, text, signatures, information within forms and legal documents, etc., by collusions over network, benefitting various industry use cases. The identifications and verifications for digital identity currently follow a centralized architecture and may or may not include deep learning technologies with more secured data-centric authentications - a gap in the overall paradigm. This leads to bottlenecks with security issues involving many concurrent processes with many concurrent parties, resulting in longer service times and a lack of trust, thus impacting the overall Quality of Service (QoS). This thesis introduces a concept called “Deep Chain Learning,” which provides a secured technique for integrating deep learning within smart contracts in a blockchain in a decentralized architecture. The smart contracts trigger execution of neural network-based models for deep learning on component images, signatures, etc., by each or some of the parties at respective nodes in a blockchain. A user-authentication mechanism allows for accessing different object components by each party in the blockchain to execute deep learning. The inference results drawn from each of these parties are written to the blockchain, and shared across all parties. The implementation of “Deep Chain Learning” uses an example driver’s license identification and verification process, as part of an auto insurance application. It also enables user access control privileges as a security measure for deep chain learning. Performance is evaluated for 3 use cases including auto insurance and healthcare applications. Results show that the distribution of inference tasks of component images among multiple parties lead to an almost linear reduction in cost, when compared to the control variable, which is a centralized, sequential mode of execution. The solution not only reduces the QoS, but with the security feature enabled, improves the overall trust in the paradigm.

Type of Resource

text

Language

eng

Owning Collections