PUKYONG

Privacy-Preserving Cross-Silo Federated Learning with a Cryptocurrency in Edge Networks

Metadata Downloads
Abstract
Cross-silo federated learning (FL) enables several devices to assemble deep learning models while keeping their private training data on the device. Rather than uploading the training data and model to the server, cross-silo FL only sends the local gradients gradually to the aggregation server back and forth. Hence, FL preserves data privacy by design. FL leverages a decentralized approach where the training data is no longer concentrated. Similarly, blockchain uses an identical manner by providing a distributed digital ledger that can cover the centralized system's flaws. Cross silo FL system lacks the proper incentive mechanism for the clients. Smart contract features can be a plausible solution as an incentive mechanism in the FL system since self-executing contracts with immutable data records are resistant to failure. Our system operates on 5G edge networks (ENs) architecture that can protect traffic between edge entry points (edge-to-edge), enabling the design of various flexible and customizable applications. In other words, we propose an intelligent system integrating blockchain technologies, 5G ENs, and FL to create an efficient and secure framework for transactions. We also extend several cryptography techniques to preserve unlinkable cross-silo FL transactions as a privacy-awareness in decentralized approaches across 5G ENs and beyond. The experimental results positively recommend that our novel models and techniques satisfy the design goals. The proposed schemes remarkably outperform the existing methods in terms of privacy-oriented cryptography in cross-silo FL empowered by blockchain technology and 5G ENs. To prove our above-mentioned research contributions, we also state a statistical significance analysis considered in the FL, blockchain, and 5G ENs inquiry.
소수 기관 대상(Cross-silo)의 연합학습 (FL)은 여러 디바이스가 딥 러닝 모델을 취합하면서 개인 교육 데이터를 디바이스에 유지할 수 있도록 한다. 교육 데이터와 모델을 서버에 업로드하기보다는 소수 기관 대상의 FL은 로컬 기울기만 왔다갔다하며 집계 서버로 서서히 전송한다. 따라서 FL은 의도적으로 데이터 프라이버시를 지킨다. FL은 교육 데이터가 더 이상 집중되지 않는 경우에서 탈중앙화된 접근방식을 활용한다. 마찬가지로, 블록체인은 중앙화된 시스템의 결함을 메꿀 수있는 탈중앙화된 디지털 원장을 제공하여 동일한 방식을 사용한다. 소수 기관 대상의 FL 시스템에는 클라인트에 대한 적절한 장려 인센티브 메커니즘이 없다. 변경 불가능한 데이터 기록이 있는 자동 실행 계약은 실패에 강하기 때문에 스마트 계약의 특성은 FL 시스템 내의 인센티브 메커니즘으로서 타당한 해결책이 될 수 있다. 우리 시스템은 에지 엔트리 포인트 (에지 투 에지) 간의 트래픽을 보호 할 수 있는 5G 에지 네트워크 (EN) 아키텍처에서 작동하여 유연하고 맞춤형인 다양한 어플리케이션을 설계 할 수 있다. 즉, 효율적이며 안전한 트랜잭션 프레임워크를 만들기 위해 블록체인 기술, 5G EN, FL을 통합한 지능형 시스템을 제안한다. 5G EN 및 그 이상에 걸쳐 연결 되지 않고 소수 기관 대상의 FL 트랜잭션를 탈중앙화된 접근방식으로 보존하기 위해 여러 암호화 기술 또한 확장한다. 실험 결과는 우리의 새로운 모델과 기술이 설계 목표를 충족한다는 것을 긍정적으로 권장한다. 제안 된 계획은 블록체인 기술과 5G EN에 의해 강화된 소수 기관 대상의 FL에서 프라이버시 지향적인 암호화 측면에서 기존 방식보다 현저하게 우수하다. 앞서 언급한 연구 기여를 증명하기 위해 우리는 FL, 블록체인, 그리고 5G EN 조사에서 고려된 통계적 유의성 분석 또한 명시한다.
Author(s)
RAHMADIKA SANDI
Issued Date
2021
Awarded Date
2021. 8
Type
Dissertation
Publisher
부경대학교
URI
https://repository.pknu.ac.kr:8443/handle/2021.oak/1123
http://pknu.dcollection.net/jsp/common/DcLoOrgPer.jsp?sItemId=200000504594
Affiliation
부경대학교 대학원
Department
대학원 정보보호학협동과정
Advisor
Kyung-Hyune Rhee
Table Of Contents
1.Introduction 1
1.1.Motivation 2
1.2.Outline of the Thesis and Contributions 3
2.Blockchain Insights and Federated Learning: An Overview 6
2.1.Blockchain: The New Technology of Trust 6
2.2.AI Enabled Blockchain Smart Contract 8
2.3.Blockchain in Federated Learning 9
2.4.Research Trends 11
3.Privacy Awareness in Decentralized Approaches 12
3.1.Blockchain-based Revenue: A Simplified Model 12
3.2.Rewarding Concerns 13
3.3.Linkability Concerns: At a First Glance 16
3.3.1.Potential Vulnerabilities and Defenses 17
3.3.2.Blockchain Dilemma 19
3.3.2.1.Block Generation and Propagation Time 20
3.3.2.2.Parameterizing the Block Size 20
4.An Extensive Empirical Evaluation of Cross-Silo FL Empowered with Decentralized Incentive Schemes 21
4.1.Introduction 21
4.2.Reliable Federated Learning with Commensurate Incentive Schemes 23
4.2.1.System Model and Analysis 23
4.2.2.Smart Contract Pipelines 26
4.2.3.Implementation and Evaluation 28
4.2.3.1.Federated Learning Model 28
4.2.3.2.Commensurate Incentive Scheme 31
4.3.Improving Data Privacy through Decentralized Predictive Model 34
4.3.1.Modeling and Operations 34
4.3.2.Simulation Results and Analysis 39
4.3.2.1.The Correctness of Comparison of Client's Devices 39
4.3.2.2.Decentralized Trading Contract 42
4.4.Conclusion 43
5.Edge Intelligence and Blockchain Empowered 5G Networks and Beyond: A State of the Art 44
5.1.Introduction 44
5.2.Fifth Generation Wireless Network (5G) 46
5.3.Practical Example: Autonomous Vehicular Networks (AVN) 47
5.4.A Concrete Scheme of Secondary Authentication in the 5G AVNs 49
5.5.Analysis of Simulation Results 51
5.6.Conclusion 54
6.An In-depth Experimental Study of Parameterizing Blockchain Core Components in a P2P Network Topology 55
6.1.Introduction 55
6.2.Fundamentals of Ring Signatures 57
6.3.CryptoNote Protocol 58
6.3.1.One-Time Use Transaction and Stealth Addresses 59
6.4.Decentralized Healthcare Management 60
6.4.1.Our Approach 60
6.4.2.The Group of Ring Signature 62
6.5.System Analysis and Design 64
6.6.Parameterizing Propagation Time 68
6.7.Conclusion 71
7.An Improved Obscuring Cross-Silo FL Techniques with Commensurate Incentive Scheme in the 5G Edge Networks and Beyond 72
7.1.Introduction 72
7.2.Blockchain-Enabled 5G Ultra-Dense Network for Cross-Silo Federated Learning 75
7.2.1.State of the Art 75
7.2.2.Cross-Silo FL Activities and Commensurate Incentive 76
7.2.3.Preparatory Experiments and Results 78
7.2.4.Rewarding Mechanism 80
7.2.5.Dynamic Authentication 83
7.3.Obscuring Cross-Silo FL Transactions 87
7.3.1.The Frameworks 87
7.3.2.Group Signatures of Clients and Terminology 89
7.3.3.Used-Model-Only Transactions 91
7.3.4.Secure Decentralized Rewarding Schemes 95
7.3.5.Research Design Limitations and Assumptions 98
7.3.6.Performance Analysis 99
7.3.7.Comparative Analysis 104
7.4.Conclusion 106
8.Conclusions and Future Work 107
8.1.Conclusions 107
8.2.Suggestions for Future Work 108
Degree
Doctor
Appears in Collections:
대학원 > 정보보호학협동과정
Authorize & License
  • Authorize공개
Files in This Item:

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.