Hardware-Coupled Bayesian Optimization for Self-Tuning Lightweight Cryptographic Parameters in Resource-Constrained Decision Support Systems

We present a self-tuning framework for lightweight cryptographic parameter optimization in resource-constrained decision support systems. The proposed method replaces static, heuristic-based parameter selection with a hardware-coupled Bayesian optimization loop that dynamically adjusts key length, block cipher rounds, and mode of operation during system operation. Our approach integrates four novel modules: a hardware-coupled cost model calibrator that builds device-specific polynomial regression models from micro-benchmarks, a transfer-learning initializer that reduces calibration overhead using a pretrained neural network, a multi-objective Bayesian optimizer that balances security strength against execution time and energy consumption, and a drift-aware feedback controller that detects concept drift in the cost landscape. The cost models are calibrated on the target device using onboard performance counters and energy monitoring circuitry, then refined through Gaussian process surrogate modeling with expected hyper volume improvement acquisition. The optimizer selects configurations via a covariance matrix adaptation evolution strategy inner loop, and the resulting measurements update the surrogate model every sixty seconds or when resource thresholds are crossed. The drift-aware controller monitors for hardware aging or environmental changes and triggers recalibration when necessary. We demonstrate that this closed-loop architecture continuously adapts to resource fluctuations and threat changes without manual intervention. The transfer-learning initializer reduces calibration time from several minutes to under thirty seconds by fine-tuning only ten percent of the micro-benchmark data. Our system achieves secure data transmission without performance degradation on typical IoT microcontrollers. The main contributions include a practical methodology for coupling empirical cost models with Bayesian optimization, a transfer learning scheme for rapid device-specific calibration, and a drift detection mechanism that ensures long-term robustness. This work bridges the gap between theoretical cryptographic parameter tuning and real-world deployment on resource-constrained hardware.