Advancing Smart Contract Security: Integrating Large Language Models with Traditional Vulnerability Detection Techniques

This thesis advances smart contract security, showcasing the effectiveness of hybrid AI-driven approaches in protecting billions of dollars within Decentralized Finance ecosystems. This thesis bridges traditional smart contract vulnerability detection with Generative AI, specifically Large Language Models (LLMs). It develops a static analysis tool to identify vulnerable smart contracts from a recent exploit. We release Detect Llama, an open-source 34B parameter model, along with fine-tuned datasets. Our work highlights improved vulnerability detection when integrating LLMs with static and dynamic analysis. Additionally, a novel transaction mutator enhances fuzz-testing through complex transaction sequences.