SNFM: Scalable Node Feature Modeling in Smart Contracts for Vulnerability Detection through Graph-based Approaches

The burgeoning security concerns of smart contracts, propelled by the significant assets they govern and the immutable nature of blockchain technology, have catalyzed the need for robust vulnerability detection methods. Traditional approaches, predominantly rule-based, are beleaguered by high false-positive rates and protracted detection times. Contemporary scholarship has ventured into deep learning techniques, yet these often neglect the nuanced semantic information embedded in contract source code. Addressing this gap, we introduce SNFM, a framework that comprehensively encapsulates the semantic and structural facets of smart contracts. SNFM harnesses both data flow and control flow information to construct a graph representation, augmented by a large language model that delves into the semantic intricacies of node representation extraction. Subsequently, the Graph Representation Learning module is employed to integrate the representations of the contract graph, facilitating the execution of vulnerability detection. Our empirical assessment, conducted on a dataset encompassing eight types of vulnerabilities, substantiates that SNFM surpasses the capabilities of existing methods.