AI-Driven Genetic Algorithms for Enhanced Numeric Feature Quantization in IoT Device Fingerprinting for Threat Detection

Detecting malicious IoT devices within a network is a crucial step in ensuring the overall security of hosts. One challenge in identifying such devices is classifying numerical features extracted from network packets. Numeric feature quantization is a technique that discretizes continuous numerical variables into a finite set of representative values. This process is essential in machine learning applications for dimensionality reduction to enhance efficiency, noise reduction to smooth out the data, and generalization to induce machine learning rules applicable to broader datasets. In this paper, we introduce SILEA2 algorithm which employs an automated approach to determining the quantization levels for numeric features using genetic algorithms (GA) and the k-means algorithm. GA helps determine each numeric feature’s optimal number of clusters. After identifying the optimal cluster count for each feature, we implement the k-means algorithm, initializing it with the cluster numbers derived from the GA process to detect the buckets in which the contiguous values are placed. When determining the fitness of each potential solution generated by GA, we utilize SILEA inductive learning algorithm as the backbone algorithm. This covering-method inductive learning algorithm efficiently extracts IF-THEN rules for classifying data points. GA ensures that each feature’s k-means clustering is more tailored and accurate. Such an automated approach allows the generation of rules that are as accuracy-gaining as possible for the dataset considered. We tested our approach to perform IoT device fingerprinting. Our approach consistently performed with the highest accuracy in 17 of 23 classifiers compared to various state-of-the-art machine learning algorithms.