Abstract: Recent progress in deep learning (DL)-based joint source-channel coding (DeepJSCC) has led to a new paradigm of semantic communications. Two salient features of DeepJSCC-based semantic communications are the exploitation of semantic-aware features directly from the source signal, and the discrete-time analog transmission (DTAT) of these features. Compared with traditional digital communications, semantic communications with DeepJSCC provide superior reconstruction performance at the receiver and graceful degradation with diminishing channel quality, but also exhibit a large peak-to-average power ratio (PAPR) in the transmitted signal. An open question has been whether the gains of DeepJSCC come from the additional freedom brought by the high-PAPR continuous-amplitude signal. In this paper, we address this question by exploring three PAPR reduction techniques in the application of image transmission. We confirm that the superior image reconstruction performance of DeepJSCC-based semantic communications can be retained while the transmitted PAPR is suppressed to an acceptable level. This observation is an important step towards the implementation of DeepJSCC in practical semantic communication systems.
Introduction
Deep learning-based joint source-channel coding (DeepJSCC) represents a significant shift in communication theory, moving from bit-level transmission to the direct transmission of semantic features. By utilizing discrete-time analog transmission (DTAT), DeepJSCC maps continuous-valued features directly to the physical channel. While this approach is highly robust to noise and offers superior image reconstruction quality, it introduces a major practical hurdle: a high Peak-to-Average Power Ratio (PAPR).
High PAPR is problematic for hardware, as it forces power amplifiers to operate in inefficient linear regions to avoid signal distortion, thereby draining batteries and limiting coverage. This research investigates whether the "semantic edge" of DeepJSCC is dependent on these high-power signal peaks and evaluates methods to suppress them for practical deployment.
System Model
The researchers developed a passband transceiver model to accurately characterize PAPR in a realistic setting. As shown in Figure 1, the DeepJSCC framework is integrated with Orthogonal Frequency Division Multiplexing (OFDM). The source signal is processed by a JSCC encoder, mapped to subcarriers, and then transformed into a time-domain signal via an Inverse Discrete Fourier Transform (IDFT). This integration with OFDM is what primarily drives the high PAPR observed in the system.
PAPR Reduction Techniques
To address the power efficiency problem, three main strategies were explored:
- Linearly Precoded OFDMA: Techniques such as LFDMA and IFDMA (similar to the single-carrier approach in 4G LTE) were used to spread signal energy and reduce peaks.
- PAPR Loss Function: A penalty term was added to the neural network's loss function during training, incentivizing the model to minimize both reconstruction error and signal peaks.
- Trainable Clipping: This novel approach implements a differentiable clipping operation within the end-to-end training loop. By simulating hardware clipping during training, the DeepJSCC decoder learns to robustly "denoise" the resulting inter-carrier interference (ICI).
Key Results and Discussion
The study's findings debunk the myth that high-PAPR signals are necessary for semantic gains. The researchers demonstrated that the signal's dynamic range can be aggressively constrained without losing the performance lead over traditional digital methods.
As illustrated in Figure 3, DeepJSCC significantly outperforms traditional digital benchmarks like BPG compression combined with LDPC coding. Even when the PAPR is suppressed, the semantic system maintains its advantage.
Key quantitative findings include:
- Massive PAPR Reduction: Unconstrained DeepJSCC exhibited a PAPR of approximately 21 dB. Through trainable clipping, this was reduced to 6 dB, which is acceptable for standard cellular hardware.
- Minimal Quality Trade-off: Reducing the PAPR by 15 dB resulted in a Peak Signal-to-Noise Ratio (PSNR) drop of only about 2-3 dB.
- Benchmark Superiority: At a 10 dB SNR and with a 6 dB PAPR constraint, DeepJSCC achieved a PSNR of ~27.4 dB, far exceeding the ~22.3 dB achieved by the best traditional digital benchmark (BPG + 16QAM + LDPC).
Conclusion
This work proves that the superior performance of DeepJSCC-based semantic communications can be preserved while meeting the power efficiency requirements of practical hardware. Trainable clipping emerged as the most effective tool, allowing the AI to learn to compensate for its own hardware-induced distortions. This is a vital step toward implementing robust, energy-efficient semantic communication in future 6G networks.
Summary updated