An Enhanced Hybrid Memetic mRMR–FOA Framework for Feature Selection in IoT Intrusion Detection Systems

Abstract

The rapid expansion of the Internet of Things (IoT) has enlarged the attack surface of connected systems, increasing their exposure to sophisticated cyber-attacks and creating a strong demand for effective Intrusion Detection Systems (IDS). High-dimensional network traffic, however, degrades both the accuracy and the efficiency of IDS classifiers. This paper proposes a hybrid feature-selection framework, Memetic mRMR-FOA, that couples the Minimum Redundancy Maximum Relevance (mRMR) filter with a memetic extension of the Fossa Optimization Algorithm (FOA). Within the framework, mRMR plays a single, well-defined role: it ranks features by an information-theoretic relevance-redundancy score and this ranking is used both to seed a high-quality initial population and to guide the memetic local search. FOA performs the global search, while a Hamming-neighbourhood local search refines elite solutions. The objective is a minimization fitness that jointly rewards classification accuracy and low redundancy while penalizing subset size. On the RT-IoT2022 benchmark, the proposed method attains the best feature-reduction ratio of 68.3 percent, selecting only 13 features at the lowest fitness value of 0.038, and improves detection accuracy to 98.43 percent. Comparative analysis against Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Grey Wolf Optimizer (GWO), standard FOA, a memetic symmetrical-uncertainty variant, and standalone mRMR confirms consistent gains in accuracy, precision, and false-alarm rate. The results indicate that the Memetic mRMR-FOA framework is an efficient and scalable feature-selection strategy for protecting resource-constrained IoT environments.