Deep learning-based vulnerability detection solutions in smart contracts: A comparative and meta-analysis of existing approaches

In the short history of smart contracts, substantial losses have occurred due to unaccounted vulnerabilities in the smart contracts loaded onto the blockchain. Vulnerabilities in smart contracts threaten the viability and confidence of blockchain technology. Machine and deep learning architectures have been increasingly proposed to assist with writing smart contracts and detecting vulnerabilities to minimize the risks of successful attacks. This paper presents the findings of deep learning vulnerability detection in smart contracts collated from selected research papers and provides an overview of the vulnerability detection architecture. This paper assessed each component of the overall architecture, individually and in combinations, to infer potential relationships in increasing detection rates in smart contract vulnerability detection tools. This study found that the area of detecting smart contract vulnerabilities is gravitating towards deep learning models, with a particular focus on combining neural networks in serial or parallel to achieve high detection results. Feature extraction of syntax and semantic information greatly determines the detection results of a model, which suggests a strong relationship between the use of source code and neural networks. Despite this relationship, given the inaccessibility and unavailability of smart contract source code, future work should focus on efficiently extracting features and context from opcodes, which are more readily available on the blockchain. Based on these findings, the study offers an in-depth discussion of the matters arising, open issues, and key aspects such as bytecode and opcode in smart contracts, failure of attention mechanisms, the rationale of hybrid models, dataset challenges, computational complexity in a large-scale blockchain environment, optimization strategies, Expunge and interdisciplinary approaches to smart contract vulnerability detection. These discussions enhance the practical relevance of the paper, highlighting the opportunities for scaling deep learning models for real-world blockchain applications.