Research On Malicious Code Identification Method Based On Deep Learning

In recent years,with the development of computer systems and network technologies,the Internet has gradually become one of the indispensable tools for people’s generative life.However,while it is convenient for people,the development of the Internet has also brought about an increase in the number and types of malicious software,and the security risks faced by the Internet have gradually attracted people’s attention.However,traditional malicious code detection and identification methods have problems such as low accuracy,high time consumption,and complex data preprocessing.As a result,the application software used in the existing Internet and industrial generation networks lacks a strong detection mechanism and process,which brings great convenience for the release and dissemination of malicious code.When the user’s system interferes with malicious code,it will cause the leakage of the user’s personal privacy and often accompanied by huge economic losses.Therefore,how to accurately and quickly detect malicious code has become a hot research issue in recent years.This paper conducts an in-depth study on the classification and identification methods of malicious codes,and analyzes the research progress of existing malicious code identification and classification methods.The deep convolutional neural network is applied to malicious code identification and classification tasks because it adopts the method of weight sharing,which makes the training parameters far less than the deep neural network,and can obtain better recognition generalization performance.However,when using deep convolutional neural networks to identify and classify malicious codes,there are often problems that code samples are of different sizes and malicious codes are lost due to pruning during the process of converting malicious codes into fingerprints.This leads to insufficient generalization ability of the trained model.For this reason,this paper proposes an image preprocessing method based on mask occlusion.During the training process of the deep neural network model,the pixels in the fingerprint image are covered by partial mask occlusion.At the same time,based on the principle of visual attention,the attention mechanism module is embedded in the convolutional neural network model to focus on the local features of the learning image.It enables the model to extract relevant features across distances,and strengthens the model’s ability to extract and learn features from data,thereby improving the recognition and classification accuracy of the model.Aiming at the problem of serious imbalance in the distribution of malicious code data,this paper proposes a random sampling and full-category sample fine-tuning model training strategy.On the basis of mask occlusion malicious code identification method,the training process is classified into two stages.First,random sampling training is adopted,which is achieved by oversampling and weight adjustment for samples with a small number of categories.Then,use the trained model to evaluate the distance measure on the training data.Select representative sample data from samples of all categories,and perform fine-tuning training on the model trained in the first stage.In this way,the model can further fit the data features to reduce the overfitting phenomenon of the deep learning method for most types of samples.Make the model fully learn the characteristics of minority class samples,improve the recognition and classification accuracy rate of the model for minority class samples,and thus improve the fine-grained recognition accuracy rate of the deep learning model for malicious samples as a whole.