We are seeing an increase in cybersecurity attacks on resource-constrained systems such as the Internet of Things (IoT) and Industrial IoT (I-IoT) devices. Recently, a new category of attacks has emerged called microarchitectural attacks. It targets hardware units of the system such as the processor or memory and is often complicated if not impossible to remediate since it imposes modifying the hardware. In default of remediation, some solutions propose to detect these attacks. Yet, most of them are not suitable for embedded systems since they are based on complex machine learning algorithms. In this paper, we propose an edge-computing security solution for attack detection that uses a local-remote machine learning implementation to find an equilibrium between accuracy and decision-making latency while addressing the memory, performance, and communication bandwidth constraints of resource-constrained systems. We demonstrate effectiveness in the detection of multiple microarchitectural attacks such as Rowhammer or cache attacks on an embedded device with an accuracy of 98.75% and a FPR near 0%. To limit the overhead on the communication bus, the proposed solution allows to locally classify as trusted 99% of the samples during normal operation and thus filtering them out.