基於聯邦學習與區塊鏈之分散式隱私保護數據交易平台

Abstract

在越來越多科技大公司爆發濫用數據醜聞後，人們開始意識到每天常用的軟體，看似免費的提供便利的服務，其實都是透過犧牲個人隱私所換來的。世界各國政府也開始意識到保障隱私的重要性，紛紛祭出嚴格的資料隱私保護法規來強化保護個人隱私，將資料所有權交還給資料擁有者。同時，付費換取資料使用權的經濟趨勢也正在興起。 這些趨勢的興起，對於運用機器學習的企業造成莫大的挑戰，使其同時面臨隱私保護法規和資料付費的挑戰，這些挑戰使得企業在日常營運中受到越來越多的限制阻礙，進而影響其企業營運。 本研究旨在建立一個分散式隱私保護的數據交易平台，透過整合聯邦學習（Federated Learning）、區塊鏈 （Blockchain）、資料品質評估（Data Quality Assessment）及容器技術 （Container Technology），提出一解決方案來解決隱私保護與數據運用的取捨挑戰。 在建構數據交易平台後，實際透過跌倒偵測（Fall Detection）具有隱私考量的案例，來測試平台的功能性、並驗證平台運用聯邦學習與資料品質評估技術的有效性。根據實驗結果顯示： （一）透過聯邦學習的方式建構跌倒預測模型能夠在 19 位資料提供者的情況下達到 95% 準確度，其表現與一般機器學習表現相當。 （二）在不同程度的資料污染情況下，資料品質評估技術皆能夠顯著提高模型準確度。 簡言之，本研究為企業端（模型需求者）與消費者端（資料提供者）提供一套具備隱私保護且可信任的數據交易平台架構，雙方皆可以在遵循隱私法規下進行隱私資料模型訓練。

After more and more data misused scandal from big technology companies break out, the people realize that the application they used daily appears to be free, but actually, all the conveniences are at the expense of their privacy. In response to public awareness of data privacy, governments worldwide started to enact the regulations to strengthen personal data privacy and give the data ownership back to individuals. Meanwhile, the trend of the pay-for-privacy economy is prevailing. The enterprises conducting machine learning as their regular business are now facing challenges from both privacy regulations and data usage payment awareness. Their regular business will become more and more restricted in the near future, affecting the bottom line profoundly. This research aims to constructs a proof of concept of a decentralized privacy-preserving data marketplace to bridge the gap between enterprises and customers. The proposed data marketplace is an agglomeration of novel technologies including Federated Learning (FL), Blockchain, Data Quality Assessment (DQA), and Container Technology. We examine the usability of the proposed application with the fall detection case and conduct the effectiveness experiments on FL and DQA. According to the results: (1) the federated learning model achieves 95% accuracy with 19 engaged subjects and is comparable to centralized settings; (2) the data quality assessment process can significantly improve the model performance under different degrees of noise. In conclusion, the proposed data marketplace provides a privacy-by-design and trustworthy platform for stakeholders to trade between models and data.