A training stop condition may be that a preset number of iterations is reached, or a trained machine learning model accomplishes a classification performance indicator. The classification performance indicator may be a classification correctness rate reaching a first preset threshold, or a classification error rate being lower than a second preset threshold.

The server may also separate a part of training samples from first training samples for use as test samples. The test samples are samples used for model calibration after model training. The trained first recognition model is calibrated with a test sample. Specifically, the test sample may be input into the trained first recognition model, and an output of the first recognition model is compared with a training label of the test sample. If a difference between the two falls within a permitted error range, the calibration of the first recognition model is completed. If the difference between the two falls outside the permitted error range, the parameters of the first recognition model are adjusted to reduce the difference between the two until completion of the calibration of the first recognition model.

The server may also establish a cost function according to an actual output and an expected output of the first recognition model, minimize the cost function using a stochastic gradient descent method, and update the model parameters of the first recognition model. The cost function may be, for example, a variance cost function or a cross-entropy cost function.

In one embodiment, the first recognition model is trained with a live facial image and a non-live facial image. The model parameters may be adjusted dynamically according to classification performance of the machine learning model, so that a training task can be completed more accurately and efficiently.