Mobile traffic classification is a topic of interest for researchers focused on improving the network capacity or for those seeking to identify potential risks to users' privacy. In recent years, traffic classification accuracy has significantly improved thanks to machine learning techniques. These techniques allow traffic identification even if it is encrypted, as in mobile networks. In this paper, we show that it is feasible to classify mobile traffic applications with high accuracy using downlink control information (DCI) messages and machine learning. The DCI messages are collected using a sniffer located near the base station. Then we extract the statistical features of the bursts and inter-burst periods of the traffic generated by mobile applications at the physical layer. This strategy uses few features and does not require a big dataset. We have tested our approach on a 4G cellular network testbed and a commercial 4G cellular network. The results show an accuracy greater than 92% and 95% for application and category classification, respectively.