Skip to main content

Table 3 Example CML methods that replace the expensive algorithmic components with their crypto-friendly versions

From: Confidential machine learning on untrusted platforms: a survey

Framework ML Algorithm Original Component Crypto-friendly Component Benefits
Mohassel and Zhang (2017) Logistic Regression, Neural Networks Sigmoid, Softmax ReLu Avoids inversion and limits expensive confidential divisions to one.
Graepel et al. (2012) LMC, Fisher’s LDA Divisions Multiplications with incorporated division factors Avoids division costs and simplifies the protocol.
Nikolaenko et al. (2013) Ridge Linear Regression LU decomposition Cholesky’s decomposition Reduces the cost complexity by half.
Nikolaenko et al. (2013) Matrix Factorization Cholesky’s Decomposition Sorting based matrix factorization Reduces the overall complexity from quadratic to within a polylogarithmic factor of the complexity in the plaintext
Sharma and Chen (2019) Boosting Decision Stumps Random Linear Classifiers Reduced number of comparisons and simplicity in learning.
Naehrig et al. (2011) Logistic Regression Exponentiation Taylor Expansion Avoids costs involved in multiple levels of multiplications.
Sharma et al. (2019) Spectral Clustering Eigen decomposition Eigen-approximation by Lanczos and Nystrom Reduces complexity of the problem from O(N3) to O(N2).