College of Information Engineering, Yangzhou University, Yangzhou 225127, China.
Research and Development Center for E-Learning, Ministry of Education, Beijing 100039, China.
Sensors (Basel). 2022 Jun 19;22(12):4621. doi: 10.3390/s22124621.
With countless devices connected to the Internet of Things, trust mechanisms are especially important. IoT devices are more deeply embedded in the privacy of people's lives, and their security issues cannot be ignored. Smart contracts backed by blockchain technology have the potential to solve these problems. Therefore, the security of smart contracts cannot be ignored. We propose a flexible and systematic hybrid model, which we call the Serial-Parallel Convolutional Bidirectional Gated Recurrent Network Model incorporating Ensemble Classifiers (SPCBIG-EC). The model showed excellent performance benefits in smart contract vulnerability detection. In addition, we propose a serial-parallel convolution (SPCNN) suitable for our hybrid model. It can extract features from the input sequence for multivariate combinations while retaining temporal structure and location information. The Ensemble Classifier is used in the classification phase of the model to enhance its robustness. In addition, we focused on six typical smart contract vulnerabilities and constructed two datasets, CESC and UCESC, for multi-task vulnerability detection in our experiments. Numerous experiments showed that SPCBIG-EC is better than most existing methods. It is worth mentioning that SPCBIG-EC can achieve F1-scores of 96.74%, 91.62%, and 95.00% for reentrancy, timestamp dependency, and infinite loop vulnerability detection.
随着无数物联网设备的连接，信任机制显得尤为重要。物联网设备更深入地嵌入到人们生活的隐私中，其安全问题不容忽视。基于区块链技术的智能合约具有解决这些问题的潜力。因此，智能合约的安全性不容忽视。我们提出了一种灵活的、系统的混合模型，称为串行-并行卷积双向门控循环网络模型与集成分类器（SPCBIG-EC）。该模型在智能合约漏洞检测方面表现出了出色的性能优势。此外，我们还提出了一种适用于我们混合模型的串行-并行卷积（SPCNN）。它可以对输入序列进行特征提取，进行多元组合，同时保留时间结构和位置信息。在模型的分类阶段使用集成分类器来增强其鲁棒性。此外，我们专注于六种典型的智能合约漏洞，并在实验中构建了两个数据集 CESC 和 UCESC，用于多任务漏洞检测。大量实验表明，SPCBIG-EC 优于大多数现有方法。值得一提的是，SPCBIG-EC 可以实现重入、时间戳依赖和无限循环漏洞检测的 F1 分数分别为 96.74%、91.62%和 95.00%。
