Blockchain smart contract vulnerability detection based on GR algorithm

Smart contract, as the core execution mechanism in blockchain technology, have their security serving as the cornerstone for the effective management of on-chain digital assets. To address the current technical bottlenecks in smart contract vulnerability detection, such as limited coverage of vulnerability categories and detection efficiency, the study was conducted. Two critical challenges were focused on: multi-category vulnerability detection and optimization of feature extraction efficiency. Methodologically, the following approaches were adopted: Five specialized datasets containing more than 3,000 annotated samples were constructed, covering mainstream vulnerability categories including reentrant vulnerabilities, timestamp dependency vulnerabilities, integer overflow vulnerabilities, transaction order dependence and transaction authorization vulnerabilities. Thus, a robust data foundation for multi-category detection was established. At the algorithmic design level, an improved GR (gated recurrent unit-random forest) algorithm model was proposed. In this model, the improved gated recurrent neural network (GRU) incorporated a decoupled attention mechanism to strengthen the ability to capture critical vulnerability features. Meanwhile, the random forest algorithm employed an information entropy optimization strategy to preserve the integrity of global features. The dual-channel processing architecture not only ensured the significant extraction of key vulnerability features, but also prevented information attenuation during deep feature transmission. Ultimately, the aim of improving both the variety and efficiency of smart contract vulnerability detection was achieved. Experimental results showed that the GR algorithm model could successfully detect five categories of smart contract vulnerabilities with an accuracy rate of 98.88%. Compared to previous algorithm models, it achieved over 3% improvement in detection efficiency and increased three detectable vulnerability categories. Thus, the feasibility and superiority of the GR algorithm model were validated.