Cache-enabled Generative Joint Source-Channel Coding for Evolving Semantic Communications

1. Problem Statement

The paper addresses critical limitations in existing learning-based Semantic Communication (SemCom) systems, particularly for image transmission in dynamic wireless environments:

• Inflexibility and Training Overhead: Current Deep Joint Source-Channel Coding (DeepJSCC) methods rely on extensive end-to-end training. These models are often fixed after training and struggle to adapt to varying channel conditions (e.g., changing Signal-to-Noise Ratios) without computationally expensive retraining.
• Redundancy in Transmission: Existing methods treat every transmission as a new event, failing to exploit the semantic redundancy across multiple transmissions of similar content (e.g., recurring facial features like eyes or mouths). This limits the potential for long-term efficiency gains.
• Robustness: Standard Generative Adversarial Network (GAN) inversion methods, while effective for compression, often lack explicit mechanisms to handle channel noise, leading to degraded reconstruction quality when transmitted directly over noisy channels.

2. Methodology

The authors propose a novel framework called CAGI-JSCC (Channel-Aware GAN Inversion-based Joint Source-Channel Coding) integrated with a Cache-enabled Dynamic Codebook (CDC).

A. Channel-Aware GAN Inversion (CAGI-JSCC)

Instead of training a new encoder/decoder pair, the system leverages a pre-trained SemanticStyleGAN model shared by both the transmitter and receiver.

• Inversion Process: The system performs "GAN inversion" to find a latent code $y$ that, when passed through the pre-trained generator $G$, reconstructs the original image $x$.
• Channel Awareness: Unlike standard inversion, the optimization process explicitly incorporates wireless channel characteristics. The objective function minimizes the difference between the original image and the image reconstructed from a noisy, power-constrained version of the latent code.
  • The process simulates the channel noise ($n$) and power normalization ($PN$) during the inversion optimization.
  • Loss Function: The optimization uses a composite loss function combining $L_1$ loss (for robustness to outliers), LPIPS (for perceptual similarity), and a task-specific loss (e.g., face recognition features).
• Training-Free: Since the generator is pre-trained and the inversion is an optimization problem solved per image (not a training of network weights), the system adapts to channel conditions without retraining the neural network.

B. Cache-enabled Dynamic Codebook (CDC)

To further reduce bandwidth, the system introduces semantic-level caching at both the transmitter and receiver.

• Disentangled Semantics: SemanticStyleGAN provides disentangled latent vectors (e.g., separate vectors for eyes, eyebrows, mouth).
• Caching Mechanism:
  • The transmitter maintains a cache of previously transmitted semantic vectors.
  • For a new image, the transmitter compares its semantic vectors against the cache using cosine similarity.
  • If a vector matches a cached one above a threshold $\gamma$, the transmitter sends only the index of that cached vector instead of the full vector.
  • The receiver retrieves the vector from its local cache using the index and reconstructs the full latent code.
• SNR-Aware Update & Eviction:
  • Upgrade: If a semantic vector is transmitted under a higher SNR (better quality) than the version currently in the cache, the cache is updated with the new, higher-quality vector.
  • Eviction: When the cache is full, an eviction policy based on a weighted combination of access timestamp (recency) and SNR (quality) is used, ensuring high-quality and frequently used vectors are retained.

C. Two-Stage Optimization

To handle the non-differentiable nature of the cache lookup (discrete indexing), the authors employ a two-stage GAN inversion strategy:

1. Stage 1: Optimize the latent code without cache constraints to explore the latent space freely.
2. Stage 2: Apply the cache operation once to fix the "hit" vectors, then refine the remaining non-cached vectors to compensate for approximation errors, using a "straight-through" estimator to allow gradient flow.

3. Key Contributions

1. Training-Free SemCom Framework: Introduction of CAGI-JSCC, which utilizes pre-trained generative models and channel-aware inversion to achieve robust transmission without end-to-end training.
2. Semantic-Level Caching: Proposal of the CDC mechanism that exploits the redundancy of disentangled semantic components across transmissions, significantly reducing bandwidth by transmitting indices instead of raw vectors.
3. SNR-Aware Cache Management: Development of a dynamic update and eviction strategy that prioritizes cache entries based on transmission quality (SNR), ensuring the cache stores the highest fidelity semantic components.
4. Robust Optimization: Integration of channel noise modeling directly into the GAN inversion objective, ensuring the generated latent codes are inherently robust to channel distortions.

4. Experimental Results

The framework was evaluated on the CelebAMask-HQ dataset (512x512 images) over an AWGN channel with SNR ranging from 0 to 5 dB.

• Perceptual Quality vs. Bandwidth:
  • The proposed system achieved a Bandwidth Compression Ratio (BCR) of 1/224 on average, and as low as 1/1024 for single images.
  • This significantly outperforms baselines like DeepJSCC and GenerativeJSCC, which typically operate at a BCR of 1/128.
  • Despite the massive compression, the system maintained comparable perceptual quality (measured by LPIPS, DISTS, FID, and visual inspection) to baselines with much higher bandwidth.
• Robustness: The system demonstrated robustness even when the transmitter had imperfect knowledge of the channel SNR (estimation errors up to 5 dB resulted in minimal performance degradation).
• Visual Quality: Visual comparisons showed that CAGI-JSCC produced sharper, more detailed images with natural textures compared to the blurry outputs of DeepJSCC and the artifacts of non-channel-aware GAN inversion.
• Cache Efficiency: As the cache filled up, the BCR decreased rapidly. The system successfully reduced overhead by reusing semantic components (e.g., specific eye shapes) across different images.

5. Significance

This work represents a paradigm shift in Semantic Communications by moving away from static, heavily trained models toward evolving, experience-based systems.

• Efficiency: It demonstrates that by leveraging the "long-tail" nature of semantic data (where common features repeat), communication overhead can be drastically reduced without sacrificing quality.
• Practicality: The "training-free" aspect makes the system highly adaptable to real-world dynamic networks where channel conditions change rapidly, eliminating the need for constant model retraining.
• Future of 6G: The integration of generative AI, semantic caching, and joint source-channel coding provides a blueprint for next-generation 6G networks that prioritize task-relevant information and resource efficiency.