Syntactic Evasion and Adversarial Detection for Language Models: A Robust Evaluation and Defense Framework. ALIGN100 6/7

This paper presents a comprehensive, reproducible framework for generating, detecting, and evaluating syntactic evasion and adversarial attacks on large language models (LLMs). The framework integrates (1) an adversarial generation suite that produces paraphrases, obfus cations, null-node insertions, and structured perturbations with mixed provenance (multiple model families and human rewrites); (2) a multi-signal detection pipeline combining calibrated uncertainty residuals, structural consistency checks, embedding drift, surface heuristics, and learned adversarial classifiers; and (3) an evaluation protocol with provenance logging, secret adversarial splits, and human adjudication. All algorithmic details, proofs, annotation protocols, reproducibility scripts, ASCII figures, and example data are included in the appendices so the paper is self-contained and ready for review. Contact: [email protected]