Convolutional Neural Network for Classification of Image-Based Malware: A Deep Learning Approach

In this digital era, people are at risk from malware in the online world. Both personal information is stolen and computer systems are harmed. Around the world, numerous programmers and information security experts are constantly developing new methods for spotting malware. Many researchers have been looking into machine learning for the classification of malware over the past few years. The current solutions are less effective for datasets with many samples and demand more computing power. Extracting features from images using existing feature extractors use up more resources. In this research, a DCNN models with pre-processing and augmenting approaches are introduced in order to classify malware grayscale photos. The Malimg dataset contains 9339 grayscale images and an investigation is conducted on this dataset. It is belonging to 25 different families and the dataset is built from binaries of malware. We presented the DCNN and hybrid models in order to develop a precise method and take into account, the performance of deep learning methodologies for characterizing the growth in newly created malware. Deep CNN is employed as a feature extractor automatically since, in comparison to the current methods, it requires less time and resources. The deep CCN model gives an accuracy of 98.3% better than the existing model. Compared to the current CNN models, the suggested deep CNN's execution time is substantially shorter. The suggested deep CNN model incorporates the SVM. Without employing Soft-max as an activation function, the features are extracted using the deep CNN model, and SVM is then used to classify the malware based on the extracted features. The proposed deep CNN model is integrated with the support vector machine. The deep CNN model is used to extract the features, and SVM is utilized to categorize the malware based on extracted features without using Softmax as an activation function. The introduction of fine-tune model of deep CNN uses the full-connected layer as input and outputs a 256 neuron features vector that has been carefully designed as input to the support vector machine. It gives 99.59% accuracy by converting binary support vector machine to multiclass SVM, it classifies sampler malware using one against another techniques.