Risk-Calibrated Semantic Transmission for Communication-Efficient Heterogeneous Collaborative Inference

This paper proposes a risk-calibrated heterogeneous collaborative inference framework that deploys a lightweight CNN at the edge and a high-capacity ViT at the server. The proposed method uses Conformal Risk Control (CRC) to calibrate the edge-side acceptance threshold with a finite-sample guarantee on the risk of incorrect local acceptance, while Adaptive Prediction Sets (APS) and Grad-CAM are combined to generate class-aware saliency maps for selective patch transmission. By transmitting only semantically informative patches when server inference is required, the proposed framework reduces communication cost while preserving inference accuracy. Experimental results on ImageNet demonstrate that the proposed method achieves 81.05% Top-1 accuracy with a communication cost of 0.26, corresponding to a 74% reduction over full-image transmission.