Salt phytoremediation from saline waters has been considered by some researchers in recent years as a natural method of desalination. For this purpose, halophytic plants are used in constructed wetlands. Due to the complexity of the phenomena governing the performance and constructed wetlands, they are often considered as black boxes for modeling. In this paper, the modeling of salt phytoremediation in constructed wetlands was performed using a multi-layer perceptron artificial neural network. The input and output layers of the networks, included five variables of Electrical Conductivity and concentration of calcium, magnesium, sodium and chloride of the influent and effluent of the studied Constructed Wetland. The data of 24 series, was randomly divided into two categories of training and test data. The networks with 1, 2 hidden layers and different number of neurons were trained and two optimal networks were selected. Selected networks were simulated using experimental data. The results indicated that the two networks were able to predict the results of the salt phytoremediation in constructed wetlands with good approximation.