A Comparative Analysis of Hybrid Deep Learning Models for Reentrancy Vulnerability Detection in Ethereum Smart Contracts

Recent research has exposed significant security vulnerabilities within smart contracts that run on blockchain. Threats, such as, reentrancy attacks, where malicious actors exploit recursive function calls in a smart contract, pose a critical threat. This led to substantial financial losses in organisations. Traditional vulnerability detection methods, largely based on static analysis, showed limitations in effectively identifying reentrancy issues, often yielding high false positive rates and missing complex execution paths. This paper analyses hybrid deep learning models for reentrancy vulnerability detection in Ethereum smart contracts, introducing a unique approach that combines semantic and syntactic feature extraction. Specifically, our approach integrates CodeBERT embeddings for deep semantic insights with pattern-based feature vectors that capture Solidity constructs that are vulnerable to reentrancy attacks. Five hybrid models are evaluated, each selected to provide insights into structural and sequential dependencies within code. Findings highlighted the novelty of using multimodal feature integration in vulnerability detection, with models like Autoencoder-LSTM and CodeBERT-Transformer Encoder achieving high accuracy of 98.3% and 98.01%, respectively, demonstrating the effectiveness of hybrid architectures for capturing complex vulnerability patterns. This comparative study advances the smart contract security field, showcasing each model’s strengths and trade-offs, and providing practical guidance for deploying deep learning-based vulnerability detection within blockchain ecosystems.