Currently, 90% of the global population relies on 4G/5G networks, with smartphones being an indispensable part of daily life. Call statistics, which are vital for billing purposes and treated as sensitive personal information, are safeguarded by legal regulations. One method of remotely obtaining call statistics involves initiating consecutive probing phone calls, which results in numerous missed calls on the recipient’s device. This paper adopts a stealthy phone call solution that utilizes the Session Initiation Protocol (SIP) vulnerabilities, enabling data collection without raising alarms for the callee. Through this approach, the paper distinguishes between calling and remaining states by analyzing the data returned from the callee. Furthermore, the paper introduces a two-level classifier to translate each call response into a state prediction, thus forming a sequence of state predictions over time to derive call statistics. To bolster the prediction accuracy of classification, two domains of knowledge, such as call state machines and typical human call behavior tendencies, are considered. This integration significantly enhances prediction accuracy to an impressive 98%. However, despite these advancements, there remain challenges. The prediction accuracy for call duration still requires improvement due to low probing frequency and occasional incorrect state predictions.