The classic logistic regression has the risk of over fitting. It can be solved by the regularization technique of the statistical learning theory. Optimization of convex loss function can ensure that the regularized risk minimization problem converges to the global optimum, but learning algorithm of convex loss function is susceptible to noise. Then T-logistic regression was proposed to amend, introducing T distribution into logistic regression. The non-convex loss function is made up for the deficiency of convex loss functions. Due to the non-convex loss function difficulty to solve, we will be logarithmic the objective function, and convex multiplicative programming is used to solver parameters. Through empirical study, it is found that T-logistic regression model has a good predictability and is tolerant to label noise.