Blind classification of LDPC codes via the Kolmogorov-Arnold transformer network

Blind identification of channel codes in a non-cooperative setting is a challenging task, especially for Low-Density Parity-Check (LDPC) codes under high bit error rates (BER). In this letter, we propose a novel blind LDPC code classification method based on a hybrid Kolmogorov–Arnold Network (KAN) and Transformer architecture. We use a Kolmogorov–Arnold Transformer (KAT) module that augments a standard self-attention block by applying a KAN-based nonlinear representation to the attention inputs, enabling robust detection of code structures even from highly corrupted sequences. The simulation results demonstrate that the proposed method can achieve superior classification accuracy and robustness compared with conventional methods.