Secrecy-aware dynamic pricing and resource allocation in MEC systems under CSI uncertainty using deep reinforcement learning

MEC-enabled IoT networks face fundamental challenges in jointly ensuring efficient resource allocation and physical-layer security (PLS), particularly under dynamic and uncertain wireless environments. In such systems, dynamic pricing and resource allocation must carefully balance secrecy guarantees and system efficiency in the presence of Channel State Information (CSI) uncertainty. To address these challenges, this paper proposes a Secrecy-aware Equilibrium-Derived Supervision guided Deep Reinforcement Learning framework for Dynamic Pricing and Resource Allocation (SeED-DPRA) in MEC systems under PLS constraints. We model the interaction between an MEC server (leader) and multiple users (followers) as a Stackelberg game, where the server determines pricing and security strategies, while users make optimal task offloading decisions. Unlike prior works that assume perfect CSI, the proposed framework explicitly accounts for CSI uncertainty and secrecy outage constraints. To ensure tractability, we derive convex robust reformulations of secrecy constraints using chance-constrained surrogates, worst-case bounding, and the S-procedure, resulting in semidefinite programming (SDP)-based equilibrium solutions. These equilibrium-derived solutions form an Equilibrium-Derived Supervision (EDS) dataset, which is leveraged to pretrain the DRL agent. To enhance adaptability under dynamic and adversarial environments, SeED-DPRA integrates EDS pretraining with a robust PPO-based fine-tuning process, augmented with secrecy-aware Lagrangian constraints and Conditional Value-at-Risk (CVaR) penalties. Simulation results show that SeED-DPRA consistently improves secrecy satisfaction, resource efficiency, and convergence speed by up to 40%, 32%, and 35%, respectively, compared to benchmark schemes under CSI uncertainty. These results confirm that the proposed framework effectively combines robust game-theoretic modeling with deep reinforcement learning to enable secure, efficient, and adaptive MEC resource allocation in practical uncertain environments.