Malware Detection Based on Multidimensional Time Distribution Features

Abstract

: Language detection models based on system calls suffer from certain false negatives and detection blind spots. Hence, the normal behavior sequences of some malware applications for a short period can become malicious behavior within a certain time window. To detect such behaviors, we extract a multidimensional time distribution feature matrix on the basis of statistical analysis. This matrix mainly includes multidimensional time distribution features, multidimensional word pair correlation features, and multidimensional word frequency distribution features. A multidimensional time distribution model based on neural networks is built to detect the overall abnormal behavior within a given time window. Experimental evaluation is conducted using the ADFA-LD dataset. Accuracy, precision, and recall are used as the measurement indicators of the model. An accuracy rate of 95.26% and a recall rate of 96.11% are achieved.