Autoencoder-Based Analog Fountain Codes for Short Packet Communications

Short packet transmission is crucial for ultra-reliable and low-latency communications (URLLC), yet designing efficient channel codes for such scenarios remains challenging due to the limited blocklength and strict reliability requirements. This paper proposes a novel end-to-end autoencoder framework for short analog fountain codes (AFC) that jointly optimizes encoder and decoder components. The framework combines a learnable AFC encoder featuring attention-enhanced weight selection with a graph neural network (GNN)-based decoder, maintaining the inherent rateless property while enabling systematic optimization. Numerical results demonstrate that our framework produces the optimized generator matrix and consistently outperforms conventional AFC schemes in terms of block error rate (BLER), approaching near-optimal error rates and realized rates at high signal-to-noise ratios (SNRs).