Designing an Efficient Memory Forensics Framework for Real-Time Malware Detection and Classification

Main Article Content

Manojkumar T. Kamble
Dr. Sridevi

Abstract

This study posits a new and effective real-time malware detection and classification scheme that utilizes memory forensics and machine learning, meeting the increasing requirement to identify obfuscated, polymorphic, and zero-day threats generally overlooked by static or network-based tools. With the CIC-MalMem-2022 dataset of more than 58,000 memory dump samples drawn from various malware families, we propose a hybrid pipeline based on a Momentum-Contrast Laplacian Autoencoder (MoCLAE), a novel embedding model that combines contrastive self-supervised learning with Laplacian graph regularization to produce high-quality, label-free feature representations. These mappings, derived solely from unstable memory structures, serve as input to supervised learners—Random Forest, Support Vector Machine, and XGBoost—to achieve precise family-level malware classification (e.g., Trojans, worms, ransomware). The work also presents improved forensic features, such as process–DLL interaction graphs and temporal delta measures like entropy change and handle count changes, to enhance interpretability and detection resilience. Experimental outcomes indicate that our framework meets near-ideal ROC AUC values over 0.999, performing better than previous models with respect to accuracy and generalization. Comparative study with newer publications (2024– 2025) affirms the newness and superior quality of this unsupervised-to-supervised methodology as a scalable and interpretable solution for real-world memoryresident malware detection.

Article Details

Section
Articles