Robust Edge Inference with Graph Neural Networks under Channel Aging

One of the key missions of sixth-generation (6G) networks is to deploy artificial intelligence (AI) at the network edge. The widely recognized paradigm, edge inference, is envisioned to enable diverse IoT applications such as autonomous driving, industrial automation, and human activity recognition. However, its efficient deployment is hindered by two major challenges: 1) the communication bottleneck from uploading high-dimensional features under limited resources, and 2) performance degradation due to outdated channel estimation, known as channel aging. This work addresses these challenges to enable robust edge inference via efficient resource allocation. In particular, we theoretically characterize the effects of channel aging on end-to-end (E2E) performance by linking inference accuracy to a surrogate metric, termed effective discriminant gain. An efficient resource allocation strategy adaptive to channel aging is realized by leveraging the permutation invariance of a graph neural network (GNN) to maximize the proposed surrogate metric. Simulation results demonstrate the superiority of the proposed framework over benchmarks that either assume perfect CSI or employ uniform resource allocation.