Adaptive Machine Learning–Driven Selective Encryption For Secure And Efficient Data Protection

Abstract

This study introduces ML-DSEA, an intelligent selective encryption framework that combines machine learning–based decision-making with deterministic cryptographic control to address the long-standing trade-off between security and computational overhead. Traditional encryption mechanisms enforce uniform protection across all data, resulting in inefficiencies in resource-constrained environments, whereas existing selective encryption techniques rely on predefined or heuristic rules that fail to adapt to diverse and dynamic data characteristics. The proposed approach employs a feature-driven learning strategy, incorporating entropy, stop-word density, and structural text attributes to estimate the required level of encryption dynamically. A comparative evaluation of multiple classifiers identifies Support Vector Machine (SVM) as the most effective model for capturing complex relationships between input features and encryption requirements. To maintain strong security guarantees, a rule-based override mechanism ensures complete encryption when the input exhibits low entropy or high predictability. Experimental validation on a dataset of 12,000 heterogeneous text samples demonstrates that the proposed method achieves 96.2% prediction accuracy, while reducing encryption latency by 28% and increasing throughput by 34% relative to the baseline DSEA method. Furthermore, security evaluation under ciphertext-only, known-plaintext, and semantic inference attack scenarios indicates improved resilience due to adaptive protection of information-rich content. Experimental results demonstrate that the proposed approach achieves up to 80% reduction in encryption workload while maintaining high classification accuracy and strong resistance against statistical and inference-based attacks. These results highlight the practical applicability of the framework for secure and efficient data transmission in real-world environments.