The invention discloses a gradient descent calculation method for protecting privacy data. The method is used in gradient function computation of machine learning including one or more data providers,a decryption service provider, and a computing resource provider,. The method is suitable for carrying out fitting calculation on a sigmoid function by utilizing a polynomial function similar to thesigmoid function contour or carrying out fitting calculation on the sigmoid function by utilizing a piecewise function similar to the sigmoid function contour. The method comprises the steps of homomorphic encryption key generation and distribution, training parameter negotiation, data encryption and summarization and gradient descent. The method is high in precision, and precision loss caused bydata processing in the calculation process is within a controllable range; the security is high, and the input data and the intermediate data can meet the semantic security requirements in the calculation process; flexibility is good, and two or more participants can participate in calculation; and the expansibility is good, and the original gradient descent can be expanded to a Newton method or abatch gradient descent.