From Rule-Driven to Data-Driven: Technological Evolution, Challenges, and Future Trends in Smart Contract Vulnerability Detection

Abstract

Smart contract vulnerability detection is undergoing a paradigm shift from rule-driven to data-driven approaches. While traditional methods relying on predefined patterns struggle with evolving attack vectors, machine learning enables novel solutions through multi-level feature learning. This study systematically traces technological evolution: Early research established foundational frameworks using syntax tree features with classical classifiers. Deep learning advancements introduced temporal models capturing execution dynamics, spatial convolutional networks decoding bytecode structures, and graph neural networks modeling cross-contract dependencies, collectively enabling precise identification of complex vulnerabilities. Semi-supervised learning and hybrid architectures further maintain detection robustness under limited labeled data. Three evolutionary trends emerge: detection scope expanding from single-contract code to multi-protocol interaction networks, feature representation transitioning from manual design to neural self-encoding, and training paradigms shifting from fully-supervised to collaborative weak supervision. Persistent challenges include unverified model interpretability, inadequate adaptation to dynamic on-chain environments, and insufficient traceability of cross-chain attack paths. Future solutions require integrating neuro-symbolic verifiability, dynamic graph networks' real-time perception, and federated knowledge-sharing mechanisms to develop adaptive, auditable next-generation detection systems.