With the growth of new technologies, biometric-based identification has been considered as an efficient method for automatic identification of individuals due to its unique nature and inability to forge it. Recently, researchers have used a combination of several different biometrics to more accurately identify people with a lower probability of error. Some of these methods use facial and fingerprint biometrics, which can become ineffective for a variety of reasons, including age and injury. As a result, choosing biometrics that are less prone to injury is an important factor. Therefore, this paper presents an identification system based on three biometrics: iris, fingerprint and face. In this method, the above biometrics are combined in two levels of feature and score, and simple and pre-trained convolutional networks are used to extract the feature from them. The results of this model on a virtual database consisting of three databases CASIA-IRIS, YaleB and FVC2000 show that the combination at the feature level gives better results due to the use of deep features. The results also indicate that the use of pre-trained network to extract features from facial biometrics, has made these biometrics more effective than the other two biometrics in accurately identifying the model.