Machine learning-driven adaptive parameter selection for homomorphic encryption in edge computing

The rapid proliferation of Internet of Things (IoT) devices and edge computing infrastructures has intensified concerns regarding data security and privacy, particularly when sensitive information is processed beyond centralized cloud environments. Homomorphic encryption (HE) offers a compelling solution by enabling computations directly on encrypted data; however, its substantial computational overhead continues to limit deployment in latency-sensitive edge applications. This paper presents a machine learning-driven framework designed to enhance the practicality of homomorphic encryption for intelligent edge environments. Our approach integrates gradient boosting optimization with the Cheon-Kim-Kim-Song (CKKS) encryption scheme to dynamically adjust cryptographic parameters in response to varying workload characteristics. We develop rigorous mathematical foundations, including formal optimization models, complexity analysis, and security proofs grounded in the Ring Learning With Errors (RLWE) assumption. We validate our framework through comprehensive experiments on the Heart Failure Prediction dataset from the UCI Machine Learning Repository, employing an expanded parameter space spanning 20 configurations across polynomial degrees from 4096 to 16,384, with security levels validated using the Albrecht-Player-Scott lattice-estimator methodology. The empirical results demonstrate substantial improvements over a conservative static baseline: latency reduction of 53.09%, throughput enhancement of 285.28%, and energy savings of 72.53%, all while maintaining computational accuracy above 0.999 and cryptographic security guarantees of at least 128 bits. Statistical analysis confirms the significance of these improvements with a p-value of [Formula: see text] and a large effect size (Cohen's [Formula: see text]). Five-fold cross-validation confirms model generalisability ([Formula: see text]). Furthermore, we provide interpretability analysis through SHAP (SHapley Additive exPlanations) values to elucidate the ML optimizer's decision-making process, addressing concerns about transparency in automated cryptographic systems. Crucially, unlike simple heuristic baselines that may sacrifice computational accuracy for speed, our ML optimizer enforces accuracy as a hard constraint, ensuring reliable results in safety-critical applications. This work offers both theoretical rigor and empirical validation, demonstrating that machine learning-enhanced homomorphic encryption can serve as an effective and secure solution for edge computing applications in healthcare, industrial IoT, and smart city domains.