Semantic-Aware Task Offloading for QoE-Driven Resource Allocation in Mobile Edge Computing

Semantic communication (SemCom) reduces communication overhead by transmitting task-relevant information instead of raw data, making it suitable for remote inference with large neural networks. We propose a semantic-aware offloading framework that combines a Vision Transformer (ViT) for semantic feature extraction with joint source-channel coding (JSCC) for noisy wireless transmission. We derive an accuracy function capturing the effects of semantic compression and channel quality on image classification tasks. We define a quality-of-experience (QoE) model considering latency, energy consumption, and accuracy. The resulting optimization problem is solved heuristically. Simulations demonstrate improved QoE, lower overhead, and better scalability than baseline approaches.