A Comparative Study of RNN-based Methods for Web Malicious Code Detection

Abstract

Malicious code can be embedded into Web applications in various ways, which will lead to frequent malicious Web attacks. In the deep learning-based Web malicious code detection methods, the effect and applicability of different RNN-based methods are unknown, which needs to be further study. Therefore, a comparative study of RNN-based methods for Web malicious code detection was conducted in this paper. Different from existing research, this paper not only analyzes and discusses the advantages and disadvantages of different RNN-based methods, including LSTM, GRU, SRU, but also utilizes Web malicious code detection as the application target to evaluate the actual performance of these methods. Experiment results show that the recall rates of GRU and SRU are 81.07% and 80.96%, respectively, which are higher than LSTM and minimalRNN. The performance of textCNN is relatively satisfactory, with scores of 90.6%, 85.54%, 87.95%, 94.4% in terms of precision, recall, F1 and AUC respectively. The comparative study displays that the performance of RNN-based Web malicious code detection methods is greatly affected by the preprocessing ways of source code.