Joint Channel and Semantic-aware Grouping for Effective Collaborative Edge Inference

TL;DR

This work tackles robust collaborative edge inference over wireless networks by formulating a joint grouping mechanism that simultaneously considers semantic relevance and wireless channel quality. It extends semantic matching with channel-conditioned attention, producing a graph of device interactions whose edges are weighted by both information utility and link state, and uses pruning to balance accuracy and communication cost. Empirical results on Imagenette using a MobileNetV3-Small backbone demonstrate that the proposed link-aware grouping outperforms local inference and disjoint semantic-only or channel-only schemes under data corruption and varying channel conditions, while reducing sidelink transmissions. The approach offers practical benefits for resilient, low-overhead edge AI in 6G and beyond, and points toward future work on query-aware compression and adaptive encoding strategies.

Abstract

We focus on collaborative edge inference over wireless, which enables multiple devices to cooperate to improve inference performance in the presence of corrupted data. Exploiting a key-query mechanism for selective information exchange (or, group formation for collaboration), we recall the effect of wireless channel impairments in feature communication. We argue and show that a disjoint approach, which only considers either the semantic relevance or channel state between devices, performs poorly, especially in harsh propagation conditions. Based on these findings, we propose a joint approach that takes into account semantic information relevance and channel states when grouping devices for collaboration, by making the general attention weights dependent of the channel information. Numerical simulations show the superiority of the joint approach against local inference on corrupted data, as well as compared to collaborative inference with disjoint decisions that either consider application or physical layer parameters when forming groups.

Figures (3)

• Figure 1: System model for the proposed collaborative inference problem: devices collect incomplete/corrupted data.
• Figure 2: Comparison of the different methods and scenarios in terms of accuracy for different query sizes.
• Figure 3: Studying the effect of the communication pruning trheshold