ACADEME: Aerial Cell-Assisted Device-Edge coinference ModEl in 6G Wireless Network

In the 6G wireless network, there may be a communication architecture with three levels of systems: user equipment (UE), non-terrestrial aerial low altitude platform (LAP), and terrestrial base station with mobile edge computing (MEC). Co-inference is the intelligent sharing of the multiple computing layers in the AI/ML model amongst the UE, LAP, and MEC. Computing, storage, and power shared between the above systems for co-inference will bring several system-level advantages. Furthermore, it optimizes the required data traffic bandwidth, energy consumption, and end-to-end latency. The available literature has analyzed the optimal split point of the AI/ML model between UE and MEC (one wireless link). However, to the best of our knowledge, there is no study on the AI/ML split model in the case of UE, LAP, and MEC architectures with two wireless links (UE to LAP and LAP to MEC). In this paper, for the first time, we propose a novel device-edge co-inference with a LAP-based aerial cell having computing power. We present a novel Aerial Cell-Assisted Device-Edge co-inference ModEl (ACADEME) algorithm that optimally assigns layers to compute to UE, LAP, and MEC according to two wireless link characteristics to minimize power consumption and latency of inference. Through mathematical modelling and simulations, we show that the proposed coinference substantially reduces latency, i.e., by 47% through the selection of the two optimal split points of the AI/ML model, one each at the UE and the LAP-based aerial cell, respectively.