AST-Level Semantic Watermarking Framework for AI-Generated Code: Robust Provenance Attribution via Structural Invariants

<title>Abstract</title> <p>The rapid proliferation of Large Language Models (LLMs) specialized in automated code generation has revolutionized software engineering while simultaneously introducing severe vulnerabilities concerning intellectual property (IP) theft, academic dishonesty, and software supply chain integrity. Existing code watermarking techniques predominantly operate during inference-time autoregressive decoding, biasing token selection via logits manipulation. However, these text-level and entropy-based watermarks exhibit catastrophic fragility when exposed to standard software maintenance workflows and semantic-preserving transformations, such as automated formatting, variable renaming, and dead-code injection. This paper proposes a novel, post-hoc, structure-aware watermarking framework that operates directly upon the Abstract Syntax Tree (AST) of the generated source code. By formalizing an invariant dictionary of deterministic, semantics-preserving mutations specifically commutative operand swapping, independent statement reordering, and control flow morphing the proposed framework embeds cryptographically verifiable signatures into the syntactic topology of the program without altering functional behavior or computational Big-O complexity. Detection is executed via a blind statistical extraction engine employing a one-proportion Z-test against a uniform null hypothesis. Experimental evaluation across standard benchmarks demonstrates that the AST-level framework achieves 100.0% detection fidelity on formatted code, 98.2% true positive rate under aggressive variable renaming, and 95.5% resilience against dead code injection, significantly outperforming state-of-the-art logits-based (UMD) and entropy-based (SWEET) baselines while maintaining a false positive rate strictly below 1.0%. This architecture establishes a scalable, model-agnostic provenance defense suitable for deployment across enterprise code assistants.</p>