Обнаружение уязвимостей умных контрактов с помощью модели, основанной на синтаксическом и семантическом совмещённом обучении

Abstract

Detection of vulnerabilities in smart contracts is of great importance for protection of digital assets. Today many researches reveal the features of the contract code using deep learning, but often use a single form of code representation. This does not allow extracting the semantics and structure of the code fully to detect various vulnerabilities. The article proposes vulnerability detection model based on fusion of syntactic and semantic features. Using TextCNN tool and graph neural networks, it is possible to extract syntactic and se-mantic representation from the abstract syntactic tree and the graph of the contract control flow. Combining the features the model increases the detection accuracy and recovery rate for five types of vulnerabilities reaching the average accuracy of 95% and the recovery rate of 91% that ensures effective detection of smart contract vulnerabilities.