Malware detection using image representation of malware data and transfer learning

Abstract

With the increased proliferation of internet-enabled mobile devices and large internet use, cybercrime incidents have grown exponentially, often leading to huge financial losses. Most cybercrimes are launched through malware attacks, phishing attacks, denial/distributed denial of attacks, looting people's money, stealing credential information for unauthorized use, personal identity thefts, etc. Timely detection of malware can avoid such damage. However, it requires an efficient and effective approach to detecting such attacks. This study attempts to devise a malware detection approach using transfer learning and machine learning algorithms. A hybrid approach has been adopted where pre-trained models VVG-16 and ResNet-50 extract hybrid feature sets from the data to be used with the machine learning algorithms. In doing so, this study contrives the Bi-model architecture where the same models are combined sequentially in the stacked form to obtain higher performance as the output of the first model is used to train the second model. With the Bi-model structure, 100% accuracy is obtained for a 25 classes problem. Performance comparison with state-of-the-art models and T-test proves the superior performance of the proposed approach.