Abstract

Goal-oriented communication (GO-COM) has recently emerged as an important concept in modern communications, owing partly to the insatiable demand for high bandwidth efficiency in edge networks and Internet-of-Things (IoT) systems. Unlike traditional communication systems focusing on packet transport and accuracy, GO-COM aims to convey information the most critical to receiver goals. To leverage the strength of emerging generative artificial intelligence (AI) models within GO-COM, this work presents an ultra-efficient GO-COM design built upon the backbone of the diffusion model. This Diff-GO+ model features high spectrum efficiency and flexible feedback control. Specifically, we embed the key information within semantic conditions and incorporate dictionary learning to derive a noise codebook for forward diffusion at the transmitter, with which a corresponding receiver model regenerates messages via denoising. Our proposed compression-friendly semantic conditions and low-dimensional codewords achieve significant communication overhead reduction and satisfactory message recovery. To control recovery quality, we introduce a "local generative feedback" (LGF) that enables the transmitter to anticipate recovery quality and ensure goal accomplishment at the receiver end. Our experimental results demonstrate that the proposed Diff-GO+ can achieve a better computation-bandwidth tradeoff with ultra-high spectrum efficiency and superior data recovery. Specifically, our Diff-GO+ can achieve 98% compression for image transmission of the Cityscape dataset.