A Transformation-Grounded Robustness Framework for AI-Assisted Smart Contract Auditing

Abstract

Smart contracts secure high-value activity across decentralized finance, digital assets, and on chain governance, yet deployed code is difficult to patch. This makes pre-deployment security assurance essential. In this work, we present a transformation-grounded robustness framework for AI-assisted smart contract auditing. Rather than reporting only baseline detector accuracy, we evaluate perturbation-response stability under semantics-preserving adversarial edits. We define an evasion-oriented threat model and explicitly cover three attack surfaces: prompt injection in Solidity comments, semantics-preserving code obfuscation, and knowledge-base poisoning of external audit context. We then con struct traceable transformed variants of vulnerable contracts and compare the proposed framework with Slither and Mythril baselines, while also benchmarking Slither, SmartCheck, Securify, and Vandal on paired original/transformed inputs. The proposed framework achieves 91.3% overall detection (1314/1440), compared with 79.9% for Slither and 85.1% for Mythril. The results also show consistent degradation in vulnerability detection under targeted transformations even when runtime behavior is preserved. These findings motivate hybrid auditing work flows that combine semantic analysis, adversarial stress testing, and verification-aware checks.