Software Vulnerability Detection Based on Anomaly-Attention

Software vulnerability is a decisive factor affecting the security of information systems. There are problems in vulnerability detection such as traditional static methods cannot cope with the detection scenarios of software without source code, and dynamic methods consume many resources are inefficient. Aiming at the potential vulnerabilities that may occur in the operation of executable programs without source code, this paper proposes a Transformer software vulnerability detection method based on Anomaly-Attention, which introduces the Anomaly-Attention mechanism in the traditional Transformer model to calculate the association differences between function call nodes to neighboring points and function call nodes to the whole sequence respectively, and then distinguish normal function call points from abnormal function call points through the association differences. The CallSeq tool is used to collect the sequence of function calls in program operation as timing data, and the fuzzy test results are used as labels to train the model in this paper. Experiments show that the method in this paper can detect vulnerabilities in unsourced executable programs with 87.73% accuracy, which is about 3% higher than the vulnerability detection accuracy of methods based on CNN, LSTM and other network models.