Mixer‐Based Neural Network for Binary Sequence Generation

ABSTRACT This paper proposes MixBinNet, a neural network‐based model that generates binary sequences with low autocorrelation sidelobes and low cross‐correlation while maintaining a balanced bit distribution. Conventional sequence design methods lack flexibility in sequence length and set size, while often failing to jointly optimize correlation and balance properties. To overcome these limitations, MixBinNet employs an embedding block for diverse initialization, intra‐ and inter‐mixing blocks that learn local and global correlation patterns and a binarization block for converting real‐valued outputs into binary sequences. In addition, we design custom loss functions that enable the joint optimization of autocorrelation sidelobes, cross‐correlation and balanced bit distributions during training. Experimental results show that MixBinNet consistently outperforms existing chaotic map‐based methods in terms of autocorrelation and cross‐correlation, while maintaining a bit distribution close to the ideal ratio of 0.5. These results demonstrate the effectiveness of MixBinNet in generating high‐quality binary sequence sets.