EPoLORE: Efficient and Privacy Preserved Logistic Regression Scheme

Logistic regression, as one of the classification method, is widely used in machine learning. Due to the complexity of training process, outsourcing the training task to a third party is a feasible choice, while the plain or direct outsourcing will inevitibaly lead to privacy leakage. To address this problem, this paper proposes an efficient privacy-preserving outsourced logistic regression (EPoLORE) scheme. In securely training the model, we design related protocols: floating-point conversion, integer multiplication, vector inner product, and activation function based on a distributed double-trap public key cryptosystem (DT-PKC), allowing the cloud server to effectively perform the integer and floating-point computations with ciphertexts of training data. In such a way, the privacy of training data is preserved and the model can obtain the accuracy approximate to that of the regular model trained in plaintext. The Security of the protocols is analyzed, thereby demonstrating that EPoLORE meets the security requirements. The corresponding experiments show the effectiveness of the proposed scheme and the comparison of model accuracy with the regular training model.