Many activities, especially in the field of cybersecurity, can be modeled using dynamic stream graphs, such as call graphs. In this work, we propose an approach aimed at detecting Advanced Persistent Threats (APTs) from their inception. Our method stands out for its ability to capture both structural and temporal information, which are key elements in distinguishing normal activities from malicious ones. To address the challenges posed by streaming processing, we rely on hashing techniques to obtain a compact representation of the data. This strategy, combined with a dynamic machine learning approach, provides rapid and incremental detection while ensuring low memory consumption. The conducted tests demonstrate the effectiveness of our method, enabling a proactive response to threats by identifying APTs at the earliest signs of activity.