Abstract: Internet backbone traffic is a complicated mix of various data flows initiated by clients via different network connections, including 3G/4G-based mobile cellular networks and wired broadband networks. Without examining application layer meta-data or inspecting into TCP/IP packet contents, existing network traffic analysis and characterization methods struggle in differentiating traffic flows from these two types of network connections. By studying the different kinds of link layer technics and wireless radio resource control (RRC) mechanisms, the traffic temporal characteristics are analyzed and formalized based on the packet delay variance. By making use of TCP/IP packet’s round-trip time (RTT) calculation, the experiments extract six significant network traffic features related to the packet delay, and apply them to train and test machine-learning classifiers to separate 3G/4G client traffic flows from broadband connection flows. These features focus on the transmission latency caused by a client’s first-hop Internet connection, and reveal the temporal variance of packet distribution from different link flows. Experiments with realistic dataset of mobile application traffic achieve a classification precision of more than 92% with effective traffic coverage and error resilience. The proposed method surpasses other related solutions also by relying on only the temporal distribution of flow packets without needing to inspect the packet content and encapsulation.