In the era of advanced networking with 5G integration, the need for efficient and scalable intrusion detection systems has become critical to securing large-scale digital infrastructures. Traditional intrusion detection approaches either analyze individual packets yielding high computational costs or rely solely on flow-based data, which can miss important sequence-level information critical to identifying interservice communications and attack behaviors. To address this, we propose a unified machine learning approach that integrates flow-based and packet-based detection using convolutional neural networks (CNNs) for advanced intrusion detection. Our method prioritizes flow-based detection for short flows as the first defense layer and selectively invokes packet-based detection for longer flows or cases deemed uncertain. Uncertain predictions from the flow-based stage are identified using a confidence threshold and re-evaluated by the packet-based system. We validate our method using a systematically generated dataset from a microservices environment alongside benchmark datasets, including CIC-IDS-2017, CIC-IDS-2018, and CREMEv2. This hybrid detection strategy yields strong performance in both accuracy and efficiency. Specifically, our approach reduces the computational cost by up to 24 × (approximately 1.38 orders of magnitude) compared to relying solely on packet-based analysis. Additionally, the model demonstrates strong generalization with detection rates of 95% and 100% for flow- and packet-based detection, respectively, even against previously unseen attacks generated through behavioral variations and command-level perturbations.