Malware Image Generation and Detection Method Using DCGANs and Transfer Learning

Cybersecurity in modern age is of utmost importance in almost every domain of economic activity. As digital activities make heavy use of multimedia a new type of cyber-threat gradually emerges: the possibility of producing and seamlessly embedding malware into digital images. Such type of malware can potentially avoid detection of typical scanners and infect the systems of either the service providers and the end-users. In this context, this study proposes and describes a complete methodology starting from the process of generation of malware-based yet realistic to the human eye images and concluding to the design of a suitable malware detector. This methodology designs and employs Deep Convolutional Generative Adversarial Networks (DCGANs) to synthetically generate two new large datasets of images: one with suspicious malware images (called Expanded Malware Images - EMI, in this study) and one with adversarial sample images of fashion products (called Fashion Adversarial Samples - FAS, in this study). The two new datasets are used for training two different Convolutional Neural Network (CNN) models using different training and configuration approaches. The first CNN (named c-CCN) follows a conventional approach for training, whereas the second one (named TL-CCN) leverages transfer learning to take advantage of the knowledge of ResNet50. Results show that the generation of malware images and adversarial samples stabilizes after 3000 iterations and produces very realistically looking images. Moreover, the TL-CNN model trained with part of the adversarial samples outperforms the other malware detector designs and produces results of high validation accuracy and minimal validation loss.