A Deep Learning Framework for Semantic Communication in Task-Oriented IoT Networks

Abstract

The exponential growth of the Internet of Things (IoT) imposes unprecedented demands on wireless network resources. Traditional communication systems, designed to achieve bit-level fidelity, are inherently inefficient for many IoT applications where the goal is to convey meaning rather than perfect data reconstruction. This paper proposes a paradigm shift from classical bit-oriented to emerging meaning-oriented communication. We introduce a novel deep learning framework for a task-oriented semantic communication system, which we name DeepSC-IoT. Our proposed system employs a convolutional autoencoder architecture, trained end-to-end to extract, compress, and transmit only the essential semantic information required for a specific task at the receiver. We evaluate our system on a visual classification task using the MNIST dataset, simulating a network of IoT cameras. Simulation results demonstrate that our semantic approach achieves a significant reduction in bandwidth usage compared to a traditional separation-based scheme (JPEG compression + LDPC channel coding + QPSK modulation) while maintaining superior task accuracy, especially in low signal-to-noise ratio (SNR) regimes. This work validates the potential of semantic communication to enable scalable and ultra-efficient massive IoT deployments.