Generative Semantic Communication for AI Model Parameters

In the future network, delivering AI models from the edge to terminals is crucial for low-latency, privacy-preserving applications. However, this process poses significant challenges in terms of transmission efficiency and the adaptability across heterogeneous terminals. To address these issues, this paper proposes a semantic-empowered model parameters transmission and generation scheme named ParaSC, which leverages the semantic characteristics of model parameters to reduce communication cost and improve system adaptability. Specifically, ParaSC designs a semantic transmitter to compress model parameters into parameter semantic symbols (PSS). Then, when a terminal requests a model, the system conducts a matching check against the existing PSSs. If matched, only the PSS is transmitted to the terminal, where a pre-deployed semantic receiver reconstructs it into model parameters. Moreover, for unmatched cases, ParaSC extends PSSs to generative purposes and proposes a model parameter generation framework based on generative adversarial network (GAN). This framework employs an optimized GAN generator to construct the model base (MB), which is defined as a synthesized semantic representation that integrates the feature extraction capabilities of multiple models. After transmitting MB to the terminal, the terminal reconstructs it into generated model parameters and adapts to the target task via lightweight fine-tuning. Experiments show that ParaSC reduces bandwidth by 10%-43.3% compared to baselines, adapts to SNR variations without cliff effects, and saves 42.8%-50% computational resources over transfer learning.