TL;DR
This paper introduces an anchor-aided two-stage training framework for multi-user semantic communication (SemCom) systems serving heterogeneous devices with diverse decoder architectures. By decoupling encoder optimization from decoder adaptation using a symmetric "anchor" decoder, the framework eliminates catastrophic forgetting caused by conflicting gradient updates. This allows a single base station encoder to efficiently and reliably serve multiple architecturally distinct users across noisy wireless channels.
What Problem Does This Paper Address?
Traditional wireless communication techniques are rapidly approaching the Shannon capacity limit. Semantic Communication (SemCom) offers a paradigm shift by transmitting only the core "meaning" or task-relevant features of data rather than raw bits, utilizing Deep Joint Source-Channel Coding (D-JSCC). However, most existing SemCom research assumes a simplistic single-user setup or multiple users sharing identical hardware and model architectures.
In real-world networks, a single base station (BS) must simultaneously serve a highly diverse population of devices—such as UAVs, mobile phones, and autonomous vehicles—each possessing vastly different computing capacities and deep learning (DL) backbones (see Figure 1).
When a single shared BS encoder is trained sequentially or jointly across these heterogeneous decoders, it suffers from catastrophic forgetting due to conflicting gradient updates. As illustrated in Figure 2, optimizing the encoder to suit one user's unique architecture inherently degrades its performance for previous users, creating a severe bottleneck for system scalability.
Methods
The authors propose an Anchor-Aided Two-Stage Training Framework that elegantly decouples encoder training from decoder adaptation:
- Stage 1: Self-Reflective Training: The BS encoder is trained end-to-end alongside an independent, perfectly symmetric "anchor" decoder. Because their architectures are mirrored, the encoder receives stable, unconflicted feedback, allowing it to fully utilize its capacity to learn a robust, generalized semantic feature space.
- Stage 2: Adaptive Training for Users: Once Stage 1 completes, the BS encoder's parameters are frozen and treated as a fixed anchor. Individual, architecturally diverse user decoders are then trained independently to align with and interpret the frozen encoder's outputs.
To validate this approach against genuine architectural diversity, the authors design and test four distinct decoder backbone families: Attention-based, Convolution-based, Residual-based (ResNet), and VGG-based models.
Datasets
- DIV2K: A high-resolution image dataset widely recognized as a benchmark for image restoration and semantic communication tasks. It was used for training and validation.
- Kodak: A standard lossless true-color image suite used exclusively to evaluate the generalization performance of the trained models.
Key Insights and Ideas
- Mathematical Proof of Gradient Conflict: The authors mathematically demonstrate (Theorem 1) that when gradients from heterogeneous user decoders are negatively aligned, joint training inherently drives up the reconstruction loss for other users, proving why catastrophic forgetting occurs in multi-user D-JSCC.
- Isolation as a Scalability Enabler: By freezing the encoder after the self-reflective stage (Theorem 2), the optimization problems for each user decoder become completely independent. This eliminates gradient interference entirely, allowing the network to scale to an arbitrary number of users without ever needing to retrain the base station.
- Architectural Proximity Influences Forgetting: A fascinating finding from their analysis shows that decoders sharing similar structural blocks (e.g., Attention and Convolution modules) experience less severe gradient conflict than fundamentally different architectures (e.g., VGG), indicating that "architectural distance" dictates the severity of semantic forgetting.
Key Results
The proposed framework was evaluated under both Additive White Gaussian Noise (AWGN) and Rayleigh fading channels against traditional "Iterative" and "Simultaneous" training benchmarks:
- Superior Metric Performance: The proposed framework achieved the highest Peak Signal-to-Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index (MS-SSIM) across all decoder types. For example, under an AWGN channel at 13 dB SNR, the Attention-based decoder reached an MS-SSIM of 0.9714, outperforming simultaneous training (0.9593).
- Robustness to Fading: Under harsh Rayleigh fading conditions at a low SNR of 1 dB, the Attention-based decoder achieved a PSNR of 26.0780 dB and an MS-SSIM of 0.8852, significantly higher than both the iterative (25.5375 dB / 0.8731) and simultaneous (25.4916 dB / 0.8715) approaches.
- Qualitative Visual Gains: Reconstructed images from the proposed framework maintained excellent color consistency and sharp pixel accuracy, successfully avoiding the block artifacts and blue-faded discoloration regions common in simultaneous training benchmarks.
Limitations and Future Work
- Modality Constraints: The current framework focuses exclusively on the image modality. Extending this cross-architecture approach to multi-modal data (combining text, speech, and video) remains an open challenge.
- Dynamic User Churn: The paper assumes a fixed and known user population during the training phase. Adapting the anchor framework to handle real-time dynamic user churn (users joining and leaving the network seamlessly) warrants further study.
- Digital System Integration: While the authors note that digital integration (such as learned constellation mappings or vector quantization) is a promising direction, it was not deeply integrated into this specific multi-user training architecture.
Summary updated