隨著一般資料保護規定(GDPR)的實施，保護機敏資料已經成為資料共享的關 鍵要素。近年來隨著聯邦式學習的出現，處理著不同組織之間共享資料的隱私 問題。然而，大多數聯邦式學習的架構，都是需要信任集中式伺服器。一旦集 中式伺服器被破壞或失去運算能力，會影響共享資料的保護。另一方面，以鼓 勵共享資料的精神，作為聯邦式學習的商業模式。 在本文中，我們提出了一個基於區塊鏈的聯邦式學習模型，應用局部差分隱私 和同態加密來保護區塊鏈上訓練結果的隱私。採用代理重新加密(PRE)演算 法，達成對每個資料請求者的客製化訪問控制。綜合以上所述，所提出的方案 享有以下優勢，包括:(1)去中心化(2)防篡改日誌(3)客製化訪問控制 (4)鼓勵資料共享。實驗結果表明，所提出模型的架構比其他方案表現得更好。 ;With the trend of enforcing general data privacy regulation (GDPR) law, protecting sensitive data has become essential in data sharing. Recently, federated learning has emerged to deal with the privacy of sharing data among different organizations. However, most of the architectures of federated learning are centralized frameworks with strong trust assumption. Once the centralized server is compromised or undependable, the protection of sharing data may break down. On the other hand, the business model for federated learning should be taken into consideration to encourage the spirit of sharing. In this paper, we proposed a blockchain-based federated learning model applies local differential privacy and homomorphic encryption to protect the privacy of training results on blockchain. The proxy re-encryption (PRE) algorithm is adopted to achieve a customized access control for each data requester. To sum up, the proposed scheme enjoys the following advantages including (1) decentralized, (2) tamper-proof log, (3) customized access control, and (4) incentive data sharing. The experimental results suggest that the architecture of our model outperformed better than other schemes.
