Malware Detection Based on Deep Learning

Abstract

The number of malicious file generation is speedy. Traditional malicious file detection is manual detection. The signatures are extracted for detection. When malicious files are generated faster than traditional malicious file detection, the latest malicious files cannot be detected, seriously threatening system security. The author used deep learning methods to detect malware (Windows malicious applications) to solve this problem. The author used the files collected by himself to make a dataset, used the image recognition method, and applied this method to identify malicious files. The accuracy of the malicious file detection model is 93.3378%. The author found that training with RGB images can save much time. Compared with training with grayscale images, the author saved 41% of the time using RGB image training, and the accuracy of training with RGB images is only 1% worse than training with grayscale images. The author used VirusTotal to analyze the files and trained the model to classify malicious files. The model accuracy of malicious file classification (68 classes) is 75.6377%, and the top 5 accuracy is 94.3747%. As the fewer classes, the accuracy is gradually improved. Using deep learning to detect malicious files can significantly reduce the detection cost, and it can detect new malicious files instantly, reducing cyber threats.