Instruction SDC Vulnerability Prediction Using Long Short-Term Memory Neural Network

Abstract

Silent Data Corruption (SDC) is one of the serious issues in soft errors and it is difficult to detect because it can cause erroneous results without any indication. In order to solve this problem, a new SDC vulnerability prediction method based on deep learning model is proposed. Our method predicts the SDC vulnerability of each instruction in the program based on the inherent and dependent features of each instruction in the Lower Level Virtual Machine (LLVM) intermediate. Firstly, the features are extracted from benchmarks by LLVM passes and feature selection is performed. Then, LLVM Based Fault Injection Tool (LLFI) is used to get SDC vulnerability labels to obtain the SDC prediction data set. Long Short-Term Memory (LSTM) neural network is applied to classification of SDC vulnerability. Finally, compared with the model based on SVM and Decision Tree, the experiment results show that the average accuracy of LSTM in classification of SDC vulnerability is 11.73% higher than SVM, and 10.74% higher than Decision Tree.