Probabilistic Semantic Filtering and Uncertainty-Aware Compression for Energy-Efficient RF Communication

We propose a novel approach to energy-efficient radio frequency (RF) communication based on uncertainty-aware semantic filtering and multitask learning. The framework utilizes a Bayesian neural network (BNN) with Monte Carlo Dropout to estimate predictive uncertainty and filter semantically redundant RF frames. This adaptive mechanism enables efficient data reduction based on confidence levels, optimizing bandwidth without sacrificing performance. The retained high-confidence frames are used for multiclass modulation classification and signal-to-noise ratio (SNR) prediction, supporting intelligent transmission under varying channel conditions while minimizing unnecessary processing. Experimental results show that entropy-based filtering achieves bandwidth savings of up to 40%, maintaining 90.80% classification accuracy, and reducing energy consumption by 10% compared to the baseline. uniform manifold approximation and projection (UMAP) visualizations confirm improved latent space separability after filtering. Energy consumption is tracked using CodeCarbon, showing a significant reduction in energy per transmitted frame. The proposed framework is lightweight, ideal for real-time inference in resource-constrained environments, and offers applications in edge artificial intelligence (AI), low-power Internet of Things (IoT), and vehicular networks (V2X), paving the way for energy-efficient communication in future 6G systems.