SELF-SUPERVISED LEARNING TO DETECT GENERAL ADVERSARIAL NETWORK MALWARE

• Main Author: Rananta Natasha, Dita
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81670
• Institution: Institut Teknologi Bandung

Cybersecurity threats are increasing with the development of malware types and variations. Effective malware detection is critical to maintaining data integrity and security. Traditionally, malware detection methods rely on definition-based signatures that are inefficient against evolving malware. Therefore, machine learning-based approaches, especially those using deep learning techniques, have become an important subject in cybersecurity research. This study presents the implementation and evaluation of Deep Convolutional Generative Adversarial Networks (DCGAN) to improve malware detection through the generation of synthetic malware samples. DCGAN, which is an extension of the Generative Adversarial Networks (GAN) architecture, has been known to be effective in generating realistic synthetic images. In this research, DCGAN was adapted to produce malware images that enable training of a more robust malware detection system. This model was trained using a comprehensive dataset containing both malware and benign samples. The primary focus is to test whether synthetic samples generated by DCGAN can be used to improve the effectiveness of malware detection systems in identifying new and previously unknown malware variants. Model evaluation shows significant improvements in malware detection capabilities. By using synthetic samples in training, the resulting malware detection system achieves detection accuracy of up to 99.5%, with very high precision and recall. This shows that the integration of synthetic samples in training datasets can enrich data variations and strengthen models against evasion techniques frequently used by modern malware.