Blind Training for Channel-Adaptive Digital Semantic Communications

TL;DR

BlindSC addresses the challenge of adapting digital semantic communication to unpredictable channels by modeling the encoder–decoder interface with parallel binary symmetric channels whose bit-flip probabilities are trainable and regularized to cover diverse environments. The approach uses continuous relaxation to enable end-to-end learning and a training-aware inference strategy that selects encoder–decoder pairs and allocates power/modulation to align actual BER with learned probabilities. Empirical results on MNIST, CIFAR-10, and STL-10 show higher PSNR/SSIM and substantial power savings compared to baselines like NECST and JPEG+LDPC, with AMPC achieving the best efficiency. This framework offers a practical route to channel-adaptive digital semantic communication with reduced training overhead and flexible deployment across varying system requirements.

Abstract

Semantic encoders and decoders for digital semantic communication (SC) often struggle to adapt to variations in unpredictable channel environments and diverse system designs. To address these challenges, this paper proposes a novel framework for training semantic encoders and decoders to enable channel-adaptive digital SC. The core idea is to use binary symmetric channel (BSC) as a universal representation of generic digital communications, eliminating the need to specify channel environments or system designs. Based on this idea, our framework employs parallel BSCs to equivalently model the relationship between the encoder's output and the decoder's input. The bit-flip probabilities of these BSCs are treated as trainable parameters during end-to-end training, with varying levels of regularization applied to address diverse requirements in practical systems. The advantage of our framework is justified by developing a training-aware communication strategy for the inference stage. This strategy makes communication bit errors align with the pre-trained bit-flip probabilities by adaptively selecting power and modulation levels based on practical requirements and channel conditions. Simulation results demonstrate that the proposed framework outperforms existing training approaches in terms of both task performance and power consumption.

Figures (10)

• Figure 1: The overall structure of BlindSC framework illustrating the approximate equivalence between blind training and actual digital communication systems.
• Figure 2: The proposed blind training strategy in BlindSC with trainable $NB$ parameters and $K$ training procedures.
• Figure 3: Convergence behavior of the average value of bit-flip probabilities ${\bm \mu}^{(k)}$ for different regularization weight $\lambda_k$.
• Figure 4: The proposed training-aware communication strategy in BlindSC with two adaptive control methods.
• Figure 5: Comparison of total power for various combinations of modulation levels.