High performance logistic regression for privacy-preserving genome analysis

Abstract Background In biomedical applications, valuable data is often split between owners who cannot openly share the because of privacy regulations and concerns. Training machine learning models on joint without violating a major technology challenge that can be addressed by combining techniques from cryptography. When collaboratively training with cryptographic technique named secure multi-party computation, price paid for keeping private an increase in computational cost runtime. A careful choice techniques, algorithmic implementation optimizations are necessity to enable practical over distributed sets. Such tailored kind Machine Learning problem at hand. Methods Our setup involves two-party computation protocols, along trusted initializer distributes correlated randomness two computing parties. We use gradient descent based algorithm logistic regression like model clipped ReLu activation function, we break down into corresponding protocols. main contributions new protocol function requires neither comparison protocols nor Yao’s garbled circuits, series engineering improve performance. Results For our largest gene expression set, train 7 billion multiplications; completes about 26.90 s local area network. The this work further optimized version which won first place Track 4 iDASH 2019 genome analysis competition. Conclusions paper, present its implementation, subprotocol securely compute function. To best knowledge, fastest existing high dimensional across