ML-Based Polymorphic Malware Detection Using Behavioural Features
Year: 2026
Type: conference-paper
Abstract: Polymorphic malware continually rewrites its own code to slip past signature-based defences, and even modern machine-learning detectors struggle when the malicious behaviour is slow, stealthy, or simply under-represented in available datasets. This work proposes a behavioural detection framework that addresses three tightly linked issues at once: the scarcity of labelled polymorphic samples, the strong overlap between polymorphic and benign runtime behaviour, and the sensitivity of detection accuracy to the chosen sampling window. The proposed approach executes samples inside an isolated sandbox with a simulated network, records system calls, API calls, and network activity at multiple time-window granularities (1, 2, and 10 seconds), and uses a Generative Adversarial Network (GAN) to synthesise additional polymorphic samples that expand the training set. The resulting models are expected to detect polymorphic variants with higher accuracy and better generalisation than single-window or non-augmented baselines, while remaining lightweight enough for real-time deployment.
Authors: Adnan Saif Alsalehi, Hussam Al Hamadi
Institution: University of Dubai
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214901700
