MATRIX: Multi-layer code watermarking based on dual-channel constrained parity-check coding

MATRIX is a multi-layer watermarking framework for code generated by Code LLMs. It models watermark encoding as constrained parity-check matrix equations and forms a dual-channel watermark through variable naming and semantics-preserving transformations. On Python code generated by multiple Code LLMs, the average watermark detection accuracy reaches 99.20%, with functional loss of only 0-0.14%, robustness against multiple attacks improved by 7.70-26.67%, and the applicable scope expanded by 2-6 times compared with existing methods.