Author(s): Sanjeevini, D Sweeya, O Ranitha and Sayad Seema

Abstract: Malicious software, sometimes known as malware, is a major concern for government agencies, corporations, and individuals concerned about online security in the modern day. Existing malware detection systems are tedious and inaccurate when it comes to detecting new viruses due to their reliance on static and dynamic evaluations of malware signatures and behavior patterns. Modern malware uses evasive tactics like metamorphosis and polymorphism to quickly change its characteristics and create a plethora of varieties. Given that most newly discovered viruses are really just updated copies of older malware, machine learning algorithms (MLAs) have become an integral part of modern malware research. This can only be achieved with extensive use of features, feature learning, and feature representation. The use of advanced MLAs, such as deep learning, allows for the complete elimination of the feature engineering stage. Even though there have been some recent studies in this area, the training data still tends to bias the algorithms' performance. We need to remove prejudice and objectively evaluate these methodologies to build new, improved strategies for efficient zero-day malware detection. To fill this void in the literature, this work investigates the use of deep learning architectures and standard MLAs for malware classification, detection, and categorization utilizing public and private datasets. The train and test segments of the experimental analysis use public and private datasets that are not discontinuous from one another but obtained across various periods. We also provide a novel approach to picture processing that is well-suited to ML and deep learning frameworks. A comprehensive experimental investigation of several methods found that deep learning architectures outperformed traditional MLAs. Taken together, our findings point to a hybrid deep learning architecture that is both scalable and successful at visual inspection for real-time malware detection. Hybrid approaches based on visualization and deep learning, whether static or dynamic, offer a new and enhanced way to successfully identify zero-day malware in big data settings.

DOI: 10.33545/26633582.2024.v6.i2b.133

Pages: 115-119 | Views: 605 | Downloads: 252

Download Full Article: Click Here