An Attention-Based Variational Autoencoder for End-to-End Wireless Communication Systems

Deep-learning autoencoders (AEs) have redefined end-to-end (E2E) wireless communication, jointly learning transmitter, channel and receiver as one network. Conventional AEs, however, transmit a deterministic latent code and impose a rigid power constraint through hard normalisation, limiting reliability at low signal-to-noise ratio (SNR). We propose an attention-augmented variational autoencoder (AT-VAE) for E2E communication. The transmitter learns the mean and log-variance of a Gaussian latent distribution; crucially, the additive channel noise is identified with the reparameterisation noise, so the physical channel performs the VAE's sampling step. Self-attention layers in the encoder and decoder capture long-range codeword dependencies, enlarging the minimum inter-codeword distance. We derive a union bound on the block-error rate (BLER), prove that maximising the evidence lower bound (ELBO) equals minimising BLER under a soft Kullback–Leibler power constraint, and extend the analysis to fading channels. Over AWGN channels, AT-VAE outperforms AE and 8/64/256-PSK baselines, with the largest gains at low SNR.