CCIHunter: Enhancing Smart Contract Code–Comment Inconsistencies Detection via Two-Stage Pre-Training

Abstract

Smart contracts are self-executing computer programs on blockchains. With the development of blockchain technology, the number of smart contracts has grown rapidly, as has the concern for their security. Regrettably, inconsistencies between the logic implemented in the code and the intentions described in the comments, known as Code–Comment Inconsistencies (CCI), are frequently present in some smart contracts. These inconsistencies can mislead readers in understanding the contract code and, in severe cases, may lead to vulnerabilities and economic losses. Existing learning-based methods are not tailored for smart contract languages, overlook the issue of insufficient context information caused by comment references and nested intentions, and rely on large-scale labeled data; whereas rule-based methods struggle to accommodate the flexibility with which developers express intentions, often resulting in false positives. To tackle the challenges posed by insufficient context information and the scarcity of labeled data, we introduce CCIHunter, a tool designed to detect CCIs in smart contracts. CCIHunter addresses the issue of insufficient context information during data modeling and incorporates a two-stage pre-training process that does not depend on labeled data to enhance its detection capabilities. Specifically, CCIHunter enhances comments based on templates and models code as a heterogeneous graph based on function calls. It utilizes CodeBERT and UniMp to generate embeddings for comments and code, respectively, and then calculates the similarity between these two embeddings. Consistency is judged by combining code embeddings, comment embeddings, and similarity scores. Notably, CCIHunter undergoes a two-stage pre-training that includes contrastive learning and mutation analysis, aiming to improve its ability to bridge the gap between code and comments and to focus on code elements at different granularities. Experimental results demonstrate that CCIHunter achieves a precision of 0.95, a recall of 0.90, and an F1 score of 0.93, outperforming existing tools.