基於同態加密的機器學習即服務的雲端外包--以 K-means 分群為例

Abstract

在機器學習 （Machine Learning, ML） 時代，人們越來越關注巨量資料在提高分析、模擬、預測和決策效率方面的經濟價值，以至於資料交易市場興起。由於 ML 需要高運算複雜度的處理過程，個人和企業傾向將雲端運算與資料市場結合使用，但眾所周知，這樣的平台在隱私保護方面存在著資料安全問題。雲端運算中最現代的隱私保護方法是全同態加密 （Fully Homomorphic Encryption, FHE）。然而，高計算成本使得傳統的 FHE 不適用於實際應用。儘管許多研究人員使用 CKKS FHE 來解決這個問題，但本研究實驗發現 CKKS FHE 中一些運算子的計算成本仍舊很高，因此本研究提出了新的安全協議來設計一種新的資料打包方法且減少使用耗時的運算子，並對本研究提出的新的安全協議外包 K-means 分群方法進行實驗與評估。結果顯示本研究方法比 SEOKC （Safe and Efficient Outsourced K-means Clustering）的分群速度更快。使用本研究的資料打包方法可以在高維度資料中獲致良好的分析效能。

In the machine learning (ML) era, people are paying more and more attention to the economic value of data in improving the efficiency of analysis, simulation, calculation, forecasting, and decision-making. It results in the rise of data markets. As ML requires high-complexity calculations, individuals and companies tend to use cloud computing with data markets. However, this platform is known to have data security issues in privacy protection. The most modern method for privacy protection in cloud computing is fully homomorphic encryption (FHE). However, the high calculation cost makes conventional FHE impractical for real-world applications. Although many researchers use CKKS FHE to resolve this problem, our experiments show that the calculation cost of some operators in CKKS FHE are still very high. In this paper, we propose new security protocols to design a new data packing method and to reduce the usage of time-consuming calculations. Then, an outsourced K-means clustering method based on these new security protocols is proposed for demonstration and evaluation. Experimental results show that our method is faster than SEOKC (Safe and Efficient Outsourced K-means Clustering). It has shown good performance in high-dimensional data analysis with our new data packing method.