<p>Silent data corruption (SDC) has become a significant threat to programs. Different techniques have been used to detect SDC instructions to protect programs. Machine learning models have been applied recently for this task. However, fault injections cause a high overhead in these approaches when determining the labels of training instructions. The labels indicate whether the instruction is an SDC instruction and reflect the SDC vulnerability of the instruction. We propose a novel Structure-Driven approach for detecting Silent Data Corruption Instructions (SD2CI). The SD2CI assesses the relationship of the SDC vulnerabilities of instructions based on the program structure. An instruction clone is proposed and combined with partial fault injection to detect SDC instructions. Experiments on several benchmarks demonstrate that the SD2CI predicts SDC instructions with 98.36% accuracy and reduces the overhead of fault injections by 30.37%. The SD2CI outperforms the graph attention network to predict SDC instructions (GATPS) based on the precision, recall, and F1-score, demonstrating its effectiveness.</p>