A Batch Verified Decentralized-AI Against Poisoning Attack In 6G Industrial CPS Environments

NCU Institutional Repository > 管理學院 > 資訊管理研究所 > 博碩士論文 > Item 987654321/93272

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/93272

題名:	A Batch Verified Decentralized-AI Against Poisoning Attack In 6G Industrial CPS Environments
作者:	詹益函;Chan, Yi-Han
貢獻者:	資訊管理學系
關鍵詞:	分散式人工智慧;區塊鏈;批次驗證;6G;CPS;投毒攻擊;隱私保護;Decentralized-AI;Blockchain;Batch Verification;6G;CPS;Poisoning Attack;Privacy-Preserving
日期:	2023-07-25
上傳時間:	2024-09-19 16:51:21 (UTC+8)
出版者:	國立中央大學
摘要:	隨著Cyber-Physical Systems (CPS) 設備的快速增長，也因此而產生了大量的數據，使得數據驅動的應用得以蓬勃發展，如智慧型系統的開發與工業4.0的實現。也因為6G網路能夠提供更快的傳輸速度和更強的連接性，進而促進了工業環境下CPS設備的部署和基於人工智慧之服務的普及性。然而傳統的分散式機器學習架構如聯邦學習，面臨了重大的安全性威脅及挑戰，像是隱私洩露和單點故障問題。此外，合作式的機器學習系統也需要強大的防禦措施來抵禦投毒攻擊之威脅。為了克服上述問題，此篇論文為工業環境下的CPS提出了一種分散式機器學習架構，利用區塊鏈技術和批次驗證，有效解決單點故障和未經授權的模型更新上傳。也提出基於Multi-KRUM演算法的模型選擇方案和利用區塊鏈實現的certificate revocation list進一步對抗了投毒攻擊。另外，此架構中的本地差分隱私機制還保證了使用者隱私，避免遭受推理攻擊。最後本篇論文中還增加了獎勵機制，使用代幣獎勵提供額外算力的參與者，從而促進合作關係以提高模型的整體準確度。在最後的實驗結果也能夠看出，本論文所提出的全面性框架增強了ICPS環境下分散式機器學習的安全性、可靠性和隱私性。;Rapid growth in Cyber-Physical Systems (CPS) devices has resulted in massive data generation, enabling the development of data-driven applications such as smart system development and Industry 4.0 realization. With the 6G network promises faster transmission speeds and stronger connectivity, fostering wider adoption of Industrial CPS devices and effective AI-based services. However, conventional distributed machine learning approaches like federated learning pose significant security challenges, such as privacy breaches and vulnerability to single points of failure. Additionally, the threat of poisoning attacks in collaborative learning systems necessitates robust defenses. To overcome these, we propose a decentralized machine learning approach for Industrial CPS that harnesses blockchain technology and batch verification, efficiently addressing single point failures and unauthorized submissions of model updates. A model selection scheme based on the Multi-KRUM algorithm and a blockchain-implemented certificate revocation list further counteract poisoning attacks. The application of local differential privacy mechanism secures client privacy against inference attacks. Finally, the use of incentive tokens serves as a motivator for clients to contribute their training results, thus promoting collaboration and improving the overall quality of the artificial intelligence model. The experimental results presented in Section Six provide compelling evidence that our comprehensive framework enhances security, reliability, and privacy in distributed machine learning within the ICPS environment.
顯示於類別:	[資訊管理研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	20	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....