Q-margin-based cost-sensitive action-value framework for decision-as-a-service in resource-constrained edge intelligence systems

초록

The proliferation of Edge IoT and fog computing infrastructures has intensified the demand for autonomous decision-making mechanisms that operate reliably under strict resource, latency, and thermal constraints. In such environments, conventional probability-based or accuracy-driven models often fail to capture the asymmetric costs and operational risks associated with safety-critical decisions. Building on the horizon-1 contextual bandit formulation of the prior intrusion detection system (IDS) baseline, this paper treats intrusion detection as a cost-asymmetric action-value problem trained through a Double-DQN pipeline in which every transition is terminal, so that the bootstrap term vanishes and the learned action value is the expected shaped reward of each decision. On this basis we propose an exponential reward shaping (ERS) mechanism with a streak counter that internalizes cost-asymmetric IDS operating points directly into the learned Q-margin, so that the deployed inference path requires no per-deployment threshold calibration. Rather than implementing a full Decision-as-a-Service (DaaS) platform, this work validates the core components of a cost-asymmetric decision framework as a building block toward DaaS in resource-constrained edge environments. Against the strongest cost-matched supervised baseline, cost-weighted cross-entropy under the same backbone, sample budget, and argmax inference, no statistically significant difference in missed attacks, false alarms, cost, or F1 is observed (all rho> 0.11 over 15 seeds); the bootstrap 95% confidence interval for the missed-attack difference is [-842, +1,012], corresponding to less than 0.07% of the attack records in either direction, while cost-weighted cross-entropy retains a small but statistically significant advantage in score-ranking quality. Within the action-value family, an ablation of the reward term against a symmetric-reward Double DQN reduces missed attacks by 36% over the same 15-seed protocol (rho < 10(-9)). The contribution is therefore a cost-asymmetric decision interface whose deployed inference path requires no per-deployment threshold calibration, together with its edge validation, rather than superior detection accuracy. The framework is validated on ARM-based Raspberry Pi 5 hardware with INT8 quantization, achieving substantial reductions in sustained-inference temperature rise and CPU utilization that support sustained operation on resource-constrained edge devices.

키워드