Adaptive ML-Driven Selective Encryption for Resource-Constrained Networks

Abstract

The rapid expansion of IoT, mobile, cloud, and edge computing infrastructures has increased the demand for lightweight encryption mechanisms capable of securing large‑scale textual communication without imposing high latency or computational overhead. Traditional full‑encryption schemes such as AES and RSA, although robust, remain unsuitable for resource‑constrained environments. Selective Encryption (SE) offers a partial alternative by transforming only critical portions of data; however, existing SE approaches rely on heuristic or deterministic rules, limiting their ability to adapt to diverse linguistic patterns. This paper introduces ML‑DSEA, a Machine‑Learning‑Driven Dynamic Selective Encryption Algorithm that integrates Support Vector Machine (SVM) prediction with the deterministic rules of the original DSEA model. ML‑DSEA extracts seven structural and linguistic features—TAC, TVC, OWcount, TVCOW, entropy, average word length, and stop‑word ratio—to estimate the optimal encryption percentage. Experimental results on a heterogeneous dataset of 12,000 samples demonstrate that SVM achieves the highest accuracy (96.2%), lowest encryption time (128.4 ms), and highest throughput (22.6 KB/s). ML‑DSEA improves security against frequency, semantic, and known‑plaintext attacks while reducing overhead by 28% compared to DSEA. These results confirm ML‑DSEA as a lightweight and adaptive encryption framework suitable for IoT, MANET, cloud, and edge devices.