Malware Detection Method Based on Adversarial Samples and a Hybrid Model

ABSTRACT Network information technology not only brings great convenience to people but also brings hidden information. With the popularization of the Internet and computer science, artificial intelligence technology ensures network security. Aiming at the problem that the accuracy of traditional malware detection is low and the recognition effect of network malware attacks is poor, a malware detection method based on adversarial samples and a hybrid model is proposed, and based on this method, network malware attacks can be accurately recognized. First, a heterogeneous graph attention network (HAN) is added on the basis of a graph attention convolutional network. Then, the dynamic attention mechanism is added to graph attention convolution, so as to enhance its learning ability of the network state. Finally, a pooling layer is added after each attention layer to hierarchically pool features, so as to increase attention to important node information. Experimental results show that microaverage accuracy and microaverage F1 score of the proposed method for malware detection are 96.92% and 97.94%, respectively. Moreover, the microaverage accuracy rates of traditional convolutional neural network, graph convolutional network, and GBDT_4 integrated classification model are 90.72%, 94.46%, and 94.03%, respectively, and their microaverage F1 scores are 90.17%, 91.72%, and 94.98%, respectively. Compared with the proposed model, their microaverage accuracy and microaverage F1 score are lower by 6.2%, 5.46%, 2.89% and 7.77%, 6.22%, 2.96%, respectively. This proves that the proposed model not only outperforms the traditional classification model in terms of classification performance and malware classification on experimental datasets but also accurately recognizes network anomalies in system testing, which has high feasibility and reliability. Therefore, the proposed method can achieve accurate detection of network malware and is effective.