Unmasking Polymorphic Malware and Its Detection Techniques in Internet of Things (IoT): A Comprehensive Review
Abstract
– Polymorphic malware changes the appearance of its code to evade detection while wisely preserving its nefarious intent. Changing encryption keys and decryption routines can make familiar threats appear different to scanners that rely on fixed byte patterns. Security monitoring on Internet of Things (IoT) devices must work with limited memory, processing power and energy, making the difficulty extend beyond recognising these mutations. This review examines polymorphic malware and its detection in IoT environments, focusing on host-based methods for Linux devices with ARM and MIPS architectures. It explains how mutation affects the evidence available to a detector and examines methods ranging from signatures and static heuristics to learned binary representations, runtime analysis, graphs, and hybrid systems. The role of large language models (LLMs) in interpreting recovered code and its behaviour is also considered. A comparison of 21 detection studies shows that network traffic benchmarks account for much of the reported evidence. These results are useful for detecting malicious communication, but they leave an important question open that, can the same evaluated methods detect previously unseen mutations of malware on the device itself? An assessment of twenty datasets and data sources identifies further limitations, particularly missing relationships between original samples and their variants, inadequate benign software coverage, and unreported analysis failures. The review sets out practical requirements for addressing these gaps, including validated mutations, testing across processor architectures, and measurement of the complete detection process. Its conceptual latency – F1 framework supports evaluating detection quality alongside delay and device resource limits. Together, these findings provide a basis for developing lightweight host-based detection whose effectiveness can be demonstrated beyond ordinary benchmark accuracy.